The whole playbook in one sitting: where AI actually pays, how to buy it without getting burned, and what to do first.
I run an AI consultancy for mid-market companies, and I built this playbook twice: once teaching it to a room of mid-market CEOs at MIT, and once running it inside real companies, including my own. Take what is useful. None of it requires a transformation office.
One claim runs through everything here, so I will state it up front. AI programs do not fail on model quality. They fail at the seams: what you measure, what you sign, what your data can answer, and what your people do on Monday. The model is the part vendors demo. The seams are the part you own. Every playbook in this library is about a seam.
The AI ROI numbers look contradictory until you see the pattern. 56 percent of CEOs report zero ROI on generative AI. Deployers of agentic systems report returns around 171 percent. 95 percent of stalled pilots have one thing in common: nobody named a metric on day one. Same companies, same tools, opposite outcomes. The two camps did not buy different technology. They measured different things.
The zero camp measures activity: seats deployed, prompts written, training hours. Vendors love activity metrics because seat licenses are how they bill. Boards do not care, because activity never shows up on an income statement. The 171 camp measures outcomes: dollars per resolution, dollars per closed deal, hours actually removed from a workflow. When the metric is an outcome, three things follow automatically: bad pilots die fast, good pilots get budget, and the vendor conversation changes shape.
The other number worth memorizing is 10-20-70. Across more than a thousand enterprise deployments, roughly 10 percent of AI value comes from the technology, 20 percent from process redesign, and 70 percent from people: fluency, champions, changed workflows. If your AI budget is 80 percent licenses and 5 percent people, you have budgeted for the zero column.
Go deeper: Why half of CEOs see zero ROI · Sell capacity, not headcount
A typical mid-market company now holds eight overlapping AI contracts nobody planned: marketing bought one, sales bought one, engineering bought one, the CRM grew one, the sponsor mandated one, and two free-tier tools quietly went to production. Meanwhile the vendors reserve the right to change the model under your workflows without telling you, train on your data, and make leaving expensive.
You do not need to evaluate the technology. You need to evaluate the proposal, and there are five clauses that belong in every AI contract: data residency, model-swap notification rights, output IP, exit and portability, and audit access. A vendor's reaction to these five is due diligence in itself. Refusing residency means your data is a training input. Refusing swap notification means your prompts are their beta test. Refusing exit terms means lock-in is the business model.
Two clarifications that a year of running this sharpened. First, "build the integration" means plumbing, not platform-building: you are customizing a car, not building one. Nobody at your size should be training models or writing harnesses; the work is wiring what exists into your workflows. Second, whether or not you win the model-swap clause, never let a production workflow ride a vendor's default model. Pin the model version your process was tested on, and treat the contract clause as the backstop, not the control.
There is also a grift pattern, and it is structural, not incidental: they closed you in one or two meetings, they are proposing a proprietary build, the deliverable is murky and the exit is opaque, they want significant payment up front, and there is no real data-work line item. Any two of those in the same proposal and you walk. Do not negotiate the flags away; the pattern is the signal.
One more thing you should know before any vendor meeting: your platform strategy was mostly decided years ago by your ERP, your CRM, and your sponsor's preferences. Companies split into a handful of stack archetypes, and the winning move is almost always to design from where your data already lives rather than from a benchmark chart.
Go deeper: The five clauses · Model-swap rights · The AI grifter test · Your stack already chose your strategy
Two mistakes kill most first deployments. The first is picking an industry-shaped problem. AI lands function by function, not industry by industry, and the functions that move first are the same everywhere: customer support, sales, engineering, back office. Waiting for an industry solution adds a year of RFPs to reach a worse answer.
The second mistake is treating an agent like software. An agent is a new hire. You would not hire a person without a job description, a manager, and a probation period, and the same discipline applies: write the six-field job spec before anyone writes a prompt (title, reports to, inputs, outputs, KPIs, and a blast radius limit: what it cannot touch without human approval). Onboard it on a workflow that already works. Run it in shadow mode alongside the human for a month. Remove approval gates only after it has been right, consistently, while someone was watching. Agents do not go rogue; they inherit. They inherit over-broad permissions, broken approval chains, and stale documents, and then they execute them faithfully at machine speed.
And before any of that: test the thesis before you scope the project. A five-day prototype against real data, wired into one live workflow, with a kill-or-scale decision on Friday, costs a rounding error. A scoped six-month pilot that stalls costs half a million dollars and a year of organizational patience. It is roughly a hundred times cheaper to be wrong on Friday.
Go deeper: Function before industry · An agent is a new hire · The five-day prototype · Anatomy of a lead qualifier · The 90-day plan
Your dashboards show outcomes: revenue, churn, pipeline. They do not show decisions: why you fired that vendor in 2023, what you ruled out, who committed to what and whether it happened. Fewer than 3 percent of companies can query their own decision history. That gap is where the durable advantage sits, because models are becoming a commodity and your proprietary context is not.
Building the brain is less mystical than the vendors make it sound, and the hard part is not storage. It is resolution: making sure the system knows that Mike, Michael S., and mike@ are one person, that "the CRM" means a specific system, that Friday's decision reversed Tuesday's. Skip that step and every answer sounds confident and cannot be trusted; more meetings produce less truth. The fix is a discipline, not a product: resolve names to identities at write time, make the system ask a human when it is unsure instead of guessing, never overwrite the original record, and keep the unanswered questions on a visible list. Feed it what you already have: meeting transcripts, email threads, decision memos. The same pipeline that files your meetings can decompose your operations into task-level units and hand you an AI roadmap your CFO can audit line by line, which is precisely how we run discovery at BuildClub.
One more habit that compounds: keep your prompts, agent instructions, and playbooks in a versioned library with review, exactly the way your engineers keep code. Five prompts in a repository today will be five hundred in eighteen months. The ones with discipline compound; the folders decay.
Go deeper: Dashboards show outcomes, the brain shows decisions · Why your AI gets confused · The resolution layer, implemented · The memory stack · From conversation to capability · A library that compounds
Because it will. Courts are already treating prompts, outputs, and agent logs as discoverable business records. 96 percent of IT organizations have deployed agents; 12 percent have centralized governance of them. If you cannot count your agents, you do not own them.
The security posture that lets you sleep is boring and architectural. No agent ever sees the whole vault: secrets live in vertical stores, one domain each, and every agent gets its own machine identity scoped to exactly what it needs, rotated on a calendar. Sensitive data gets a clean zone: contain it, transform it there, and let only the transformed output leave. And every agent has a tail: observability, evaluations, fallbacks, and retention are line items you pay on schedule or pay in incidents. The prototype cost five thousand dollars; running it responsibly costs that monthly, and the CTO already knows.
Also, a quiet succession problem is sitting in your org chart right now: your best people's AI work lives in personal accounts. The day one of them resigns, four custom assistants and a year of institutional context walk out the door. Claim the accounts while everyone is friendly.
Go deeper: The secrets playbook · Every prompt is a document · Every agent has a tail · The clean zone · When the AI walks out the door
Every workforce splits three ways on AI, roughly 10 percent adopters, 60 percent curious, 30 percent refusers, and each tier needs a different play. Promote and fund the adopters; they are already building the workflows that actually ship, usually in spite of policy. Invest fluency in the curious; they decide whether the whole thing reaches escape velocity, and fluency is training and champions, not license counts. Reassign the refusers first and exit second, and learn to name the saboteur pattern: endless process objections, committee demands, data-quality doubts raised at every result. It reads as engaged. It functions as obstruction.
And while you modernize the workflows, capture the judgment. Every company has one person who knows where the bodies are buried and why the vendor relationship survived 2019. That knowledge is a balance-sheet item that appears on no balance sheet, and it compounds or evaporates depending on whether you run the campfire: a weekly recorded hour where the keeper tells the stories, and a process that turns stories into rules and rules into systems your AI can use.
Go deeper: Three tiers, three plays · Manage the change or it manages you · The campfire protocol
One last thing, and it is the closest thing I have to a warning. The failure mode I now see most is not the failed pilot. It is the company that gets one process working, declares AI done, and stops. Most teams are using a small fraction of what their standard tools can already do, and the compounding returns live past that point: looping back, adding more sophisticated techniques, taking the next process. This playbook is a flywheel, not a checklist.
If you need the map, that is what a Phase Zero assessment produces: a task-level view of where AI pays in your business, with numbers your CFO can audit. And if you have already done that discovery yourselves, bring it: a Pilot Build starts from your map, distills it, and ships the first automation. Two doors, same destination.
Need the map? A Phase Zero assessment builds it, task by task, with numbers your CFO can audit. Already have the map? A Pilot Build starts from your discovery and ships the first automation.
See how we work together Start with Phase Zero