Home / How I Can Help / Enterprise AI Consulting
Service · Strategy through delivery
Enterprise AI consulting, from outcome to production
Most enterprise AI programmes do not fail on the model. They fail on picking the wrong first workflow, granting the wrong amount of autonomy, and discovering the governance answer after the incident rather than before it.
The first mistake is usually the workflow. Teams pick the most visible process rather than the one where an agent has a clear boundary, a checkable outcome and a cheap failure. Visible and tractable are not the same property, and starting with the visible one means the programme is judged on its hardest case.
The second is autonomy treated as a switch. Blanket-on produces the irreversible action nobody sanctioned. Blanket-off produces a queue of approvals that get rubber-stamped inside a week, which is the same thing with a paper trail. Autonomy is a per-task setting, and it should start where reversibility and blast radius say it should, then move up as the system earns it.
The third is the number. A programme reports cost per contact, or per interaction, or per token, and then somebody senior asks what it costs to resolve a customer issue and the two figures are not comparable. Gartner expects that by 2027, 40% of enterprises will demote or decommission autonomous AI agents because of governance gaps found only after a production incident. Most of that is these three things.
This is for you if
- → You have several AI pilots and no agreement on which should go to production first.
- → A governance, risk or security function is asking questions the delivery team cannot answer yet.
- → You need a cost model per business outcome rather than per API call.
- → You want somebody who has been on both sides: the architecture and the board conversation.
Where I am not the right fit
- × You want an AI strategy deck for a leadership offsite. Useful things, not what I do.
- × You want a vendor shortlist compiled. I will examine a shortlist, not assemble one for a fee.
- × You want the programme validated. If the first workflow is the wrong one, that is what you will hear.
- × The blocker is that two executives disagree. No external opinion resolves that, and pretending otherwise wastes your money.
- 01
Pick the first workflow properly
Score candidate workflows on whether the outcome is checkable, whether failure is reversible, whether the data exists, and who owns the result. The right first workflow is often unglamorous, which is exactly why it works.
- 02
Set autonomy per task
Reversibility and blast radius set the starting level. Proven reliability and detectability earn each move up. Written down per task, with the conditions for turning it back down, because autonomy lapses and someone has to notice.
- 03
Answer the governance questions early
Security review, data access, deployment ownership and cost ceilings are the four that no compensating control covers. Each gets a named person who answers it and an artifact that proves the answer, before go-live rather than after.
- 04
Build the cost model that survives a challenge
Every cost line, not just inference, divided by a business outcome rather than an interaction count. Written so a finance partner can reproduce it, because a number that cannot be reproduced does not survive its first review.
- 05
Make it observable before it is live
Traces over logs, with the decision path visible, so that when something goes wrong in production the team can see what the system did rather than infer it. Retrofitting this is how incidents stay unexplained.
- → A ranked shortlist of workflows with the reasoning written down, so the choice survives a change of sponsor.
- → An autonomy map: per task, the level granted, what earned it, and what would take it back down.
- → A gate list separating genuinely blocking checks from compensating controls, each with an owner and an artifact.
- → A cost model per business outcome, with assumptions published so finance can reproduce it.
- → An observability plan that names the signals to capture before the system goes live.
- → A memo written to be forwarded, and a readout with both the leadership and the engineers.
What to Automate First
How to choose the first workflow, and why the most visible one is usually the wrong answer.
Read it → ArticleTurn Up AI Autonomy Gradually, Like a Dial, Not an On/Off Switch
Five notches, and the two axes that decide where a task starts and what earns a move up.
Read it → ArticleAI Agent Observability Is Not Just Logs
The signal layers you need in place before the first production incident, not after it.
Read it →Frequently asked
Quick answers
- Where should an enterprise start with agentic AI?
- With a workflow where the outcome is checkable, the failure is reversible, the data already exists, and one person owns the result. Not with the most visible process, which is the common instinct and which means the programme gets judged on its hardest case. Pick something unglamorous with a clear boundary, get it into production properly, and use what you learn to choose the second one.
- How much autonomy should we give an AI agent?
- It depends on the task, not the agent, and it should change over time. Two things set the starting point: how reversible the action is, and how large the blast radius is if it goes wrong. Two things earn a move up: demonstrated reliability on that specific task, and whether you would detect a failure. Write it down per task, along with the conditions that would take it back down, because autonomy granted once tends to stay granted.
- What does enterprise AI consulting cost?
- The Readiness Review is fixed price and public: around $750 at the pilot rate and around $1,500 at the standard rate, for a 90-minute working session, a memo and a readout. Larger work is scoped per project with a fixed deliverable, a fixed price and a defined number of revisions. There is no public price ladder for build work, because scoping it before anyone has seen the system produces a number that is wrong.
- Do you work with our security and governance teams directly?
- Yes, and it goes better when that happens early. Four of the gates, security review, data access, deployment ownership and cost ceilings, cannot be compensated for by anything else, and each needs a named person who answers it and an artifact that proves the answer. Bringing those functions in at the design stage is much cheaper than discovering their position two weeks before a planned go-live.
Start a conversation
Bring the programme, not just the pilot.
A 30-minute call. Tell me what is running, what is stuck, and what leadership has been promised. If there is a fit we will scope it tightly in writing.