May Habib

The through-line

The argument May Habib returns to in every setting is that the gap in enterprise AI is not model capability, it is everything around the model. "Individual productivity does not rewire your organization for agentic productivity" [1], and so the vast majority of companies have "not seen an iota of business impact despite millions and tens of millions spent" [1]. Her diagnosis of why is consistent: executives tell people to change things with AI and then hand them copilots, "and you're not going to get the wholesale reinvention that actually drives impact" [3]. What produces impact instead is end-to-end workflow rewiring, org design change pushed top-down, and a vendor intimate enough with the customer to co-create it [1][12]. Writer is positioned as an enterprise agentic platform for regulated industries, with brand, compliance and governance built in [1][12], and its founding conviction, held since 2020, was that governance of language, whether brand, legal or regulatory, would be the heart of the platform [1].

The second constant is her sense that she has been working on the same problem for a decade. Language technology started as a personal preoccupation, moved through machine translation, and became an intelligence business "because language now is the interface to intelligence and soon to be super intelligence" [6]. What has shifted over time is the scope of ambition: from AI-assisted writing to AI writing to "who knows what's beyond" [12], and now to a claim that the model itself becomes the system of record and the customer's own IP [1][6].

On why enterprises spend millions and get nothing

Her account of failed AI programmes is structural, not technical. She describes a top 50 US financial institution that wanted to email 11 million people about tax changes in a new bill and simply could not, because a legacy SaaS stack made it impossible to work out fast enough what had changed and who would be affected: "it was such a missed opportunity," and initiatives like that are exactly what drives ROI [3]. The blocker is organisational. Silos have to be broken to change workflows end to end, and "you've got to really kind of break those silos" is among the reasons companies have not seen returns [3]. She reports the same message from every executive conversation: "we've got adoption. I forced people to use this stuff, but I have no results to show for it" [12]. Pressure on those executives is intense, with Fortune 500 boards meeting on weekends about AI [12].

She is blunt that she is willing to talk about the thing others avoid: "No one wants to talk about ROI, right? All the cool boys aren't talking about ROI" [12]. She also distinguishes two milestones that get conflated. Writer can show agentic capability running on real workflows in 30 to 60 days, and that gets presented to boards hundreds of times, but production and scale are different things, and the 80 or 100 million of savings from turning off agencies or rewiring an organisation "takes time because it's about people's jobs and livelihoods and roles and responsibilities and leadership paths," which makes it "painstakingly slow in the enterprise" compared with the coding market [12]. Politics decides how ambitious a project can be. Where a CIO, CMO and CEO are splitting the baby on strategy, or where there is a sunk-cost allegiance to an internal company GPT that cost 20 to 40 million and is accumulating tech debt as soon as it launches, the work shrinks to making one person's patch bright [12]. Where organisational health is good, as at Edward Jones, Vanguard and Northwestern Mutual, customers tell her they moved from level two of their plan to level ten [12]. Buyer psychology adds inertia: "the adage of, you know, you don't get fired for buying IBM applies to the guy who's at 5 million a month on Anthropic overages," and until the first person gets fired, no one will [1].

On the labs, and selling to customers rather than at them

Her sharpest competitive framing is about posture. The labs "come into enterprises with a ton of fanfare, um you know, like they're celebrities charging you millions of dollars to use their products, and then they're out of there," so a flashy demo never comes to life because they are "talking at customers rather than to customers" [1]. She notes their research and model teams are separated from product, that they interact little with customers, spend little on implementation, and have created separate deploycos to handle it [1]. Put more directly: "You've got Anthropic and OpenAI sales people who walk in like heroes, get a contract, and leave to literally never be seen again," which leaves enormous room for startups like hers [12]. When she is told Microsoft or an enterprise-styled Anthropic could build the shared-workflow functionality, her answer is that it is not even on their roadmap, "They don't even know it's a problem, right? Which is the problem" [12].

The alternative she describes is a single team where "everything from the model to the product touch each other" so the LLM is responsive to enterprise needs [1], plus formal structures for learning across customers: a CIO council, a CISO council, a CMO council, and industry-specific ones [12]. She frames the company's own origins the same way, as "a customer service organization, a client outcomes organization that had a research lab. Right? Not not the other way around" [1], and traces the instinct to growing up in a family business, listening to her father haggle with suppliers at his tool and die company and doing the bookkeeping [1]. The result, she argues, is a level of co-creation that has no precedent: "I don't think anyone has had this level of intimacy with the customer ever. Like, not Palantir, not Salesforce" [12]. Her own habit reinforces it, since she says she cannot use software without filing Jira tickets and product design requests [12].

On being full stack, and building your own models

Writer trained its own models when the conventional advice was not to, and she treats that as a contrarian bet that paid off: people said the frontier models are getting more powerful and cheaper, why not just build on them, and building their own "has really proven to be what the Enterprise needs" [2]. The platform is interoperable, with an LLM gateway built into the product so a customer like Nvidia can run its own models and anything in Bedrock can be used inside Writer [3], yet more than 90 percent of volume still runs through Palmyra, "the workhorse of Agentic um in terms of accuracy, speed, um uh how we use the context window, how we tie into our native retrieval um uh features and just overall cost and performance" [6][12]. Her reasons for training in-house are enterprise readiness, scale, profitability and inference speed, whether through synthetic data generation or fine-tuning for the specific tasks people do in the product, which beats "literally every API call being to a third party you know generic massive model" [3]. Asked whether it is cost or efficiency, she says both, and adds that if you can do it, "you'd be crazy not to" [3].

Synthetic data is the technique that made this affordable. Writer trained a separate LLM that takes real factual data resembling its use cases and converts it into structured form for clean training, an approach the company had been working on for years and one that serves customers who are not GPU rich [4]. She rejects the echo-chamber and hallucination objection by drawing a quality line: "IF YOU WERE CREATING GOBBLEDY GOOK SYNTHETIC DATA, I WOULD AGREE," but the point is to build "DATA MEANT FOR DATA" rather than reusing data meant for humans [4]. She also argues this is a genuinely different path rather than a stopgap, since "LARGER DATA SETS ARE HITTING THEIR CEILING AND THE FUTURE BELONGS TO THIS MORE PRECISION TRAINING APPROACH AND THE ARCHITECTURE INNOVATION ON THE TRANSFORMER ITSELF" [4]. On cost at the customer end, Writer's scaffolding and preference-setting mean 30 to 50 percent fewer tokens per interaction, because workflows are set in advance and the model is not rethinking how to do something: "You're not setting out a new set of plates every time someone comes over for dinner with Writer" [12].

The strategic case for full stack is control of the roadmap. Because the company is intimate with the technology, it can productise ahead of scaling results showing up, planning features for enterprise memory, context engineering and the LLM as system of record while research is still working on them [6]. That matters commercially because enterprises need to know what is coming, not what shipped, in order to decide what goes behind a beta flag and what enablement to build [6]. She also names the risk of dependence in unusually concrete terms: "it is really hard to stomach the idea that like we could be letting Vanguard down because OpenAI quantized the model and didn't tell anybody," or being reliant on a third party that begins to compete with you [6]. She contrasts this with the industry's lag pattern, where a new model arrives and six months later all the agent builders appear, just as the rag companies appeared a generation before [6]. Being one of only six or seven labs training from scratch also shapes hiring: researchers join a much smaller lab but get a narrower set of problems that reach production in weeks, and a roughly hundred-person EPD team, three times bigger than a decade ago, is twenty times more productive [6].

On governance, brand as code, and encoding a company's values

Governance is the product, not a compliance wrapper. The first transformer problem the company found interesting was brand governance, using a transformer to change language to make it more brand compliant or more suitable for translation [1], and that system of governance, "whether it's brand, whether it's compliance, whether it's legal," has been the heart of the platform ever since [1]. Brands like Skittles at Mars and Burt's Bees at Clorox are managed in Writer, with guidelines and brand books accessed by agents but never trained into the underlying model, because that is customer IP [1]. Her North Star for output is that "what you get from Writer is true to your essence, true to your brand, compliant, and, you know, easy to put out in the world because it checks all the boxes" [1].

The technical expression of this is what she calls the new CMS, "not the content management system. It is the context management system," together with tooling to track and audit an agent's decision-making and an ontology constructed agentically from how users actually work, so the nth interaction is better than the first [3]. She claims this is a structural advantage: "our competition doesn't have systems that are getting smarter the more people use them in your side of your company" [3]. A context management system also lets an enterprise encode not just knowledge, process and know-how that LLMs lack, but its value system, which is her answer to whether neutral or objective enterprise AI is realistic across cultures [3]. In practice that means customers building an SEC audit agent and a FINRA agent to review everything produced, a brand agent, and in one case an ethics agent [3]. The commercial payoff is consistency: ask Copilot or ChatGPT the same question fifty different ways and you get 48 or 49 slightly different answers, which "doesn't pass the bar" in regulated industries [5]. Much of the work, she says, is making "explicit what the organization knows implicitly but doesn't apply consistently" [5]. Writer wins where use cases demand high precision, quality, consistency and reliability, and she sums up the positioning as "Writer is for the stuff you have to get right, that you cannot afford to get wrong" [5][12].

She applies the same logic to authenticity and human relationships. In financial products or advice, "you simply cannot email somebody with LLM voice" because the client knows you [12], and nobody will trade productivity against brand dilution or risk [12]. To keep volume from eroding care, Writer ships policies that let sales ops, sales enablement or marketing ops define how connectors may be used, for example requiring a human check before an email sends and disallowing automated flows for one-to-one messaging, "the feature set that you build when you actually talk to customers versus talk at customers" [12]. Her earlier work also included mitigating bias, copyright infringement and hallucination in LLM use, and Writer's choice of graph databases over vector databases for generative AI [10].

On bias and who builds the models

She sees a "very disappointing fall-off" in attention to trust and bias, not only in LLM training but in technology at large [3]. Her case for caring is commercial as much as ethical: if an ELF Cosmetics user asks Writer to change something about a product image and the system changes the race of the person holding it, "you're not going to trust our product," so testing has to be done in the shoes of customers who want to build inclusive companies and inspire brand trust [3]. Asked whether bias is now a data problem, a governance problem or a power problem, she answers that it is a prioritisation problem: the evals exist, the frameworks exist, much of it is public, so the question is "is there a Jira epic for it or not?" [3]. Diversity, she says, starts at the top with a diverse team, tracked internally with data and once published externally [3]. She has been explicit that Writer's roughly half women and non-binary composition was "a very conscious Choice," adding "I'm in charge now and I don't want to work with only men as much as I love them," after years in male-dominated manufacturing, investment banking, private equity and tech where she says she never felt othered [2].

The equity concern extends to who benefits from the technology. Writer surveys customers on productivity gains and finds the middle of the distribution at 50 to 60 percent more productive, but the tails matter more: some people are 200 percent more productive and some are not touching it at all [2]. AI could be "10 maybe hundred times more in equity creating or Equity creating right depending on how we shape it," and without deliberate attention to user experience and enablement the result could be "a really dystopian future where most jobs are done by AI employees and folks who already would have been top of the food chain" gain all the leverage [2]. She has also argued that AI can replace the tasks people hate without eliminating jobs, and can help users with neurodiversity issues such as ADHD [10].

On the builder class, power users, and the people who block change

Adoption, in her telling, is a question of character rather than title. The partners worth having are "iconoclastic" thinkers, open and curious about a different way of doing things, and you can tell the difference immediately: "you can tell when someone's there cuz they've been asked to be there," moving the way someone who has been told to move moves, versus someone who cannot wait to get out of bed and work on AI [1]. Across five years she has seen "no correlation in title, even in seniority, even in background. It is a character and a personality," typically creative, risk-taking entrepreneurs who found themselves in enterprise environments [1]. A Writer survey of about 3,000 C-suite executives found that companies rating themselves furthest ahead on AI had cultivated an elite group of AI power users [1]. But the power law holds inside that elite: power users are single digits at most customers versus over 50 percent inside Writer, the vast majority of a company does nothing, and a small minority is actively sabotaging [1]. The playbook is to mind meld with the executives and iconoclasts, empower the builders and power users, and then win over the skeptics by showing them how critical they still are and how much more has opened up for their contribution [1].

Product design follows from this. Writer's North Star is that "a normal person needs to be able to use this," with the infrastructure handled behind the scenes, and what emerges is "a builder class within a team," a non-technical power user who thinks in systematic ways and builds the workflows everyone else benefits from [12]. The reward for being that person can be dramatic: in an executive business review, the CEO of a top 50 professional services company noticed one name attached to agents used hundreds of times across the company, called him into the meeting on Teams, and asked whether he wanted to be head of AI [3].

On careers, hierarchy and what to study

The thing she thinks is under-discussed is that "career ladders are dead" [3]. You can no longer progress by further and further specialisation or gain authority by seniority in title; hierarchies and silos are flattening, and stature will come from the amount of impact you have, which means adding adjacent functions to your scope rather than narrowing [3]. Writer's best power users illustrate it: "they don't think about their jobs as task execution. They see their jobs now as orch designing and orchestrating systems that get tasks done," which by definition means owning more adjacent things [3]. The same logic underlies her view that sales and marketing should operate as one team, since the technology compresses the time to get ideas to market [1] and shared go-to-market workflows demand "a radical rethinking of their org design," the biggest single blocker to scale [12]. She frames the wider shift as execution going "from scarce and expensive to abundant and on demand, which means human value is going to acrue to those who can orchestrate execution versus those who literally execute" [5]. Organisations that succeed will bring employees along, tell them how they will be successful, and "really forcibly flatten the hierarchy" [3].

Her advice to someone entering the workforce is short. Study critical thinking, and you do not have to be a lawyer to take pre-law and learn it [3]. Then join very collaborative, non-siloed teams, because it is shockingly common for one team or business unit not to talk to another, and it is very hard for a junior person to have impact with AI in that environment [3].

On what happens next: voice, agentic databases, and the rewrite of the stack

She declines three-year predictions but is confident about six months. First, most AI use moves to voice and mobile. Writer Channels, launched internally and to a closed group of customers as "a enterprise grade open claw," has her talking to her phone constantly, including in the elevator, and the point is invoking your agent inside conversations with colleagues [1]. For a salesperson in the field, speaking an insight to an agent can feed back into the company in a way that influences the next campaign, which is the real-time intelligence loop she expects to arrive quickly [1]. She says she now talks to her phone more than to everybody else combined [12].

Second, the sales and marketing tech stack gets completely rewritten. CIOs who want leverage and see what agents do in sales and marketing are asking whether they still need "this incredibly expensive and what now feels like truly archaic workflow processes" [1]. Writer's answer is operating agents, an agentically constructed database that is self-learning and self-improving, with the model as system of record and that record being customer IP, so companies stop working around legacy marketing automation or CRM and adopt "a truly agentic operating rhythm" [1]. Her underlying claim is that tools determine workflows, that enterprise work has been siloed and slow because of what the tools allowed, and that fiefdoms people thought were safe will be disrupted [1]. Further out, she expects LLMs to become proactive systems with memory that understand your routine and prompt you: "Hey May, you didn't ask me this week about XYZ, right?" [5]. She has also flagged an interest in brain-computer interfaces as a future application area [5]. The same rewrite is already visible in what customers build: healthcare plans that spent tens of millions on outsourced development for medical record summarisation are now rewriting that tooling themselves, because "it is now 10,000 times easier to build software" [3], and she describes the broader shift as a complete rewrite of software with enormous last-mile work still to do [4].

On the bubble, and where the pain lands

Asked whether there is a bubble, she separates the question from whether AI is real. One way to see it: the big labs are each valued at more than Salesforce plus Adobe plus Databricks plus Snowflake combined, which is hard to wrap your head around given how hard those companies are working to bring AI to their own customers [3]. The bull case is that if customers build their own software rather than buying it off the shelf, and if agentic work covers everything that was not possible before, the software market could be ten times bigger [3]. Her honest conclusion is split: "absolutely, you know, if you're looking at spending that is happening now, it is very hard to figure out how the math squares," and if anyone gets hurt she expects it to be the bigger players [3]. She is similarly clear-eyed about the noise around the category. People assume AI is a handful of companies controlling everything, and she counters that it is easy to miss how many gaps exist in the enterprise and how little impact has landed outside coding [1]. Asked how Writer competes with Microsoft, OpenAI and Anthropic, her answer is "quite easily, actually," because of the size of the gap they have left [1].

On product market fit that has to be won again and again

She rejects the idea of product market fit as an achievement you bank. Writer is "in a constant state of reestablishing product market fit" because the market is "radically noisy," noisier than any market that has ever existed by orders of magnitude, with everyone carrying ChatGPT on their phone, bouncing between vibe coding tools, and watching demos on Twitter and LinkedIn [6]. Enterprises see 40 vendor pitches a month; a pharma CIO came away from a two-day innovation event convinced fifteen startups were ready for prime time, and will meet a completely different fifteen next quarter [6]. Companies reach 100 million quickly and it is "easy come easy go" [6]. Her guidance to founders is therefore that fit "is going to be established and lost reestablished and lost again all in the span of 12 months potentially," and "just because they're still paying you doesn't mean you still have product market fit" [6]. The discipline is to be honest: take a customer's stated love of the product, then run their use case in Copilot yourself and understand the delta, which puts constant pressure on product and customer success to go deeper, better and faster [6].

The counterweight is depth of relationship and specificity. Against giants, she goes in with "a lot more specificity and actually a lot more referenceability," able to say Writer has transformed medical writing or regulatory submissions end to end for a life sciences company's peers [3]. She also names what is hard about being smaller: leaders at OpenAI are not constantly re-emphasising to employees how they will win, and she and other founders of single-digit-billion-valuation companies going up against hyperscalers have to reinspire people daily by showing customer impact [3]. Differentiation got harder in the agentic era when everyone went horizontal and used the same words, to the point that a prospect asked a Writer rep to redo sales materials without using any of the words Copilot uses [12]. Her response was to work backwards from deliverables, showing that a Monday morning briefing for 4,000 pharma salespeople has to look the same, be audited the same, and plug into Outlook, the CRM, and spreadsheets where there is no CRM, all built centrally, with interfaces that are not "hey, how can I help you," which she calls "the new hamburger menu" [12].

On risk, discomfort and the origins of the company

Her decision rule is to lean into what frightens her, a habit she traces to being the eldest child of an immigrant family: "you're kind of the Navigator and you are often in uncomfortable situation," so "over time I've gotten comfortable being uncomfortable being scared" [2]. She was born in a rural Lebanese village of about ten houses on the Syrian border, and the family left in 1990 after her mother told the Canadian Ambassador in Damascus she wanted to take her children from the war [2]. She is the oldest of eight, and connects her appetite for risk to childhood freedom, roaming from the age of six and coming home when the streetlights came on, which she contrasts with the way she raises her own children now [5]. She is equally candid about the counterweight: "there is a bit of kind of immigrant paranoia and fear that can hold me back from taking big risks," so she works against it deliberately by surrounding herself with people who do not think that way [2]. She describes her natural mode as incremental unless forced into a corner, with an explicit 80/20 split between scaling what works and betting on what might not [2], and says "I actually wish we failed more because I want more at bats to be able to tell people that that's okay" [2]. The company's risk-taking is not gut-driven but signal-driven: one customer here, two there, a prospect somewhere, "find signal and noise marry that to conviction and experience and just take risk" [2].

She insists the risk runs both ways. Because the technology can credibly address a company's biggest problem, customers taking that on face hallucination, copyright, legal and regulatory exposure, so "it takes risk taking on both sides to make generative AI work" [2]. That is why she thinks the vendor-customer relationship in generative AI is unlike any other, collaborative and transparent, with Writer only a few steps ahead of its customers [2]. She still remembers bristling the first time she was called a vendor [12]. Going enterprise rather than credit-card self-serve was itself the harder choice, taken because big problems and big companies mean bigger impact faster [2]. Platform was another contrarian call: rather than letting customers stitch together a DB here, a prompt-chaining tool there and a rails library elsewhere and end up with non-production-ready proofs of concept, Writer put the tools in one platform and offered deployment in customers' own environments, breaking the startup rule of forcing everyone into multitenant [2].

The origin story runs through language. She studied economics and near eastern languages and civilizations, worked at the Crimson redlining copy against the style guide, and wrote the first Crimson article on Facebook [5]. Being the eldest in an immigrant clan made her "a translator and not just of language but of culture" [5], and her founding conviction was that "the language you were born speaking shouldn't impact the kind of life you end up leading," because someone who cannot speak their community's language fluently "literally can't join the modern economy" [2]. She left Lehman Brothers and Mubadala, where she covered semiconductors and repeatedly tried to get the fund to invest in a chip startup called Nvidia, because "if I didn't leave um and really start, you know, start something, I never would" [5]. With Wasim, whose Arabic tokenisation NLP repos on GitHub she had seen, she built statistical machine translation, sold translation services to generate training data, and shipped a localization product that opened PRs with localized strings [6]. The pivot came when an early model produced a haiku on a slide at an all hands: "it was like what the is this" [5]. Shutting down revenue was painful, investors thought they should sell the business, and the transition took about 18 months of sunsetting revenue and rebuilding the team, "the eating glass pain of transitioning customers and humans right out of the business" [5][6][10]. She now calls that the reason Writer had a couple of quiet years before the hype cycle and a five-year head start serving regulated enterprises [5], and reflects that after ChatGPT she briefly wondered whether they should have gone consumer before concluding pure-play enterprise was right [5].

Takeaways

• The failure mode in enterprise AI is handing people copilots and productivity tools when the returns come from end-to-end workflow rewiring, org redesign and top-down leadership: "individual productivity does not rewire your organization for agentic productivity" [1][3].
• Production and scale are different milestones. Writer can show working agentic workflows in 30 to 60 days, but the 80 to 100 million in savings takes far longer because it touches jobs, roles and leadership paths [12].
• Own your models where you can. Palmyra handles over 90 percent of Writer's LLM calls for accuracy, speed, context handling, retrieval and cost, while an LLM gateway keeps the platform interoperable with Bedrock models and customers' own [3][6][12].
• Synthetic data is a precision-training strategy, not a stopgap: a separate LLM converts real factual data into structured training data, on the view that larger data sets are hitting their ceiling [4].
• Governance is the product. Brand books and guidelines are accessed by agents but never trained into the underlying model, and the "context management system" encodes a company's knowledge, processes and value system, spawning SEC, FINRA, brand and even ethics agents [1][3].
• Bias is a prioritisation problem more than a data problem, because "the evals are there, the frameworks are there," so the real question is "is there a Jira epic for it or not?" [3].
• Adoption follows character, not title: single-digit percentages of a customer's staff are power users, most do nothing, a minority sabotages, and the win comes from empowering iconoclasts and then converting skeptics by showing what has opened up for them [1].
• Career ladders are dead, and value accrues to people who orchestrate systems rather than execute tasks; for new entrants, study critical thinking and join collaborative, non-siloed teams [3][5].
• Product market fit in generative AI is temporary and must be re-won: "just because they're still paying you doesn't mean you still have product market fit" [6].