You Can Automate Insight, Not Absolution
How to buy AI now: build in-house, rent the logo later
“Insight is now a commodity. Absolution remains a luxury good.”
- Stuart Winter-Tear
If you only have 5 minutes
Consultancies sold the frontier and delivered theatre. Buyers are now building in-house where it touches the ledger, then bring the logos back later for defensible scale. The WSJ reporting shows the pattern: partners “learning on our dime,” year-long GenAI engagements ended and insourced, and a likely second wave in four to five years when the work becomes reliable and less differentiating. Buy accordingly. Build state-changing workflows internally now, prove them with receipts, then rent the logo later for assurance, rollout, and cover.
The moment: theatre punctured, value unresolved
Consultancies wagered billions that they would be essential to the AI boom. In the first wave, executives hit a delivery gap. Partners could talk about the future and spin up proofs of concept, then stalled on scale. “We love our partners, but oftentimes they are learning on our dime,” said one CIDO. That is not hostility. It is a value problem.
Bristol-Myers Squibb is the clean signal. They ended a year-long GenAI engagement aimed at doctor education content and moved the work inside. Their CTO and CIO added that for hands-on tools like Gemini CLI or Claude Code, a partner might have no more practical experience than a kid in college. Buying “frontier expertise” collapses when you already live at the frontier.
Spend is still rising, results are uneven. Consulting revenue tied to GenAI is up sharply year over year. Some firms report new bookings in the hundreds of millions, although momentum has softened compared with previous quarters. Simple story: money is flowing; certainty is not.
Why chat windows fail and workflows win
You do not move a ledger by admiring work from a chat window. You move it by changing state in the flow that carries value. That means agents that act under rules, receipts that prove what happened, and rails that constrain the system where the work lives. Many internal teams are better placed to make real use cases work today than external generalists who are still catching up.
The absolution economy
Facts are cheap, knowledge is abundant. What remains scarce in large organisations is permission to act and insulation if it goes wrong. That is why the logo matters when AI gets boring and standardised. A credible third party turns proofs into policy, results into rollouts, and risk into something a board can defend. Analysts expect the profitable second wave to arrive as technologies become reliable, predictable, and less tied to competitive advantage.
So split the market in your head.
Frontier means invention, uncertainty, and speed, and it belongs inside the business.
Mature means playbooks, accreditation, and rollout, and it is where you rent the logo.
What leaders should actually buy now
1) One value stream, end to end
Choose a flow where money moves, for example fraud checks, KYC, underwriting triage, reconciliations, claims, or collections. Define the start, the state change, and the receipt. If it does not touch the system of record, it will not touch the ledger.
2) Agents with keys, not copilots with opinions
Your agent must change state under policy. Give it permissions, escalation rules, and a hard stop when confidence drops. No free-floating assistant above spaghetti.
3) Evaluation and audit in the loop
Decide ground truth, sampling, and pass rates before anyone ships. Put evidence capture in the workflow so receipts write themselves. Your risk partner should be able to read the ledger of decisions without a side spreadsheet.
4) Governance that ships
Controls live where the action lives. Kill-switches, delegation limits, and model change logs belong in the system, not in a binder.
5) Capability, not dependency
Upskill a core internal team to operate, tune, and extend the agent. Use external help for capacity spikes, not for the core brain of the work.
What to buy later from the big firms
When you have receipts in one stream:
External validation and board-grade narrative. Independent assessment of controls, model risk, and business impact, which is permission to move budget.
Repeatable rollout. Standards, process libraries, procurement patterns, and vendor diligence at scale.
Regulatory posture. Mapping of operating controls to sector guidance and auditor expectations.
This is the profitable second wave. The product is not invention. It is legibility and assurance.
How to avoid “learning on our dime”
Use this in RFPs and steering groups.
Proof of scale, not demo flair
Two production references where the system changed state in a regulated flow, with quantified deltas to cost, error, or cycle time. Names can be redacted, numbers cannot.
Hands that can type
Name the practitioners who will do the work and cite their last two hands-on builds. Senior faces, junior hands is a smell.
Receipts by design
Show where the receipt is written in the workflow. If evidence capture is a separate task, it will be skipped.
Governance in the system
Point to the kill-switch, the approval thresholds, and the change log. Do not accept PowerPoint governance.
Commercial terms that match uncertainty
Pay for outcomes or capability. Avoid open-ended discovery that funds someone else’s learning curve.
Metrics that matter
Set these before anyone opens a laptop.
Throughput per 1,000 items in the target flow, before and after
Error rate by class for decisions the agent makes or proposes
Median end-to-end cycle time for the flow, not just the agent step
Escalation rate and time to human resolution
Cost per decision including compute, licence, and rework
Control health: number of kill-switch events, model version drift, audit completeness
If the proposal cannot show how each metric is measured in-flow, it is theatre.
The executive call
Right now, the frontier belongs to teams who touch the work and can change it. That is why buyers are pausing big consulting programmes and building in-house first, while keeping budget for logos later when the task shifts from invention to assurance. Do both: in order. Build one working agent inside a real flow, prove it with receipts, wrap it with evaluation and controls, then bring in a logo to turn a working pattern into policy, rollout, and board-safe scale.
Industry receipts
Partners “learning on our dime” reported by senior technology leaders
Year-long GenAI projects ended and insourced by buyers who live at the frontier
GenAI consulting spend up strongly year over year, while quarter-to-quarter momentum softened for some firms
Analysts expect firms to profit more in four to five years when work is reliable, predictable, and less differentiating
Big claim tension: “double your share price in five years” sits beside “few clients are at full potential today”
Coming next: my book for executives
From Hype to Hard ROI in the Age of AI is a short, practical field guide for leaders who want to leave theatre behind and move a ledger. It shows how to select one value stream, design agents that change state under policy, bake in evaluation and audit, and ship governance inside the workflow. It includes buyer checklists, acceptance gates, and board-grade narratives that help you secure permission to act.
What you will walk away with: a field-tested way to move one value stream, secure permission to scale it, and defend it at the board.
If you want receipts and rails instead of theatre, this book will give you both.
Thanks for the thoughtful write-up, Stuart. I guess I have several thoughts after reading this post...
1) It feels like the same commentary/criticism can be applied to consultants during ANY groundbreaking technology shift. I would bet similar playbooks were run in the early days of mobile broadband, cloud, etc... However, it feels like things are moving MUCH faster in the AI world - with much higher financial stakes given the insane amount of capital being thrown at this challenge.
2) To reap the true benefits, business leaders need to approach the integration of AI like any large "transformation project"... However, like many many large business transformation projects - a non-trivial percentage of them fail because of poor planning, unrealistic expectations, ill defined requirements, etc. And in the case of AI, there's so much FOMO going around that people seem to be jumping into the deep end without any semblance of a plan at all.
Overall, I love the pragmatic roadmap that you lay out here. Hopefully more business leaders will follow this guidance as we all try to figure out how best to integrate AI into our workflows.
I appreciate your contributions and thought leadership!
I love this series you started - guessing excerpts from your book? Clear guidance, actionable insights, no hype. Always left wanting to read more.