AI strategy consulting: what it should deliver
Most AI strategy engagements produce a list of opportunities and a maturity score. Both are easy to write and neither tells you what to build on Monday. The version worth paying for hands you an order, a shape decision, and one number to be judged against.
Below: the five things the engagement should deliver and the test for each, the four questions that decide its shape, a three-week sequence, and four situations where you should not buy one at all.
Published August 22, 2026· 10-minute read · Written for multi-location and mid-market operators
Five deliverables, and how to test each
A ranked build sequence, not a list of opportunities
An opportunity list is the easy half and it is where most engagements stop. What you need is an order, with the dependency reason attached to each position: this one goes third because it reads from the record the second one produces.
The test: Ask what goes first and why the thing after it cannot go first. A strategy that could be executed in any order is a list.
The source-system inventory
Every system holding a customer record, a location record, or a piece of brand truth — including the spreadsheets nobody admits to. Almost every AI plan founders here, because the plan assumed one source of truth and the operation has nine.
The test: Export your location or customer list from every system that holds one and count the disagreements. That number is the real starting position.
A shape decision
Whether you need a tool, an orchestration, or an operating change. These are different purchases with different budgets and different failure modes, and conflating them is the single most expensive error in this category.
The test: Ask the consultant to argue for the cheaper shape. If they cannot, they have one product.
The governance design, at the same time
Thresholds, routing tiers, and what gets logged — designed alongside the build rather than added after. A control retrofitted onto a running system is a control nobody designed the system around.
The test: Ask what the auto-publish threshold will be and in what units. Vagueness here means governance is a later phase, which means it is a later invoice.
A number to be judged against
One measurable outcome with a baseline and a date. Not "AI maturity" and not a capability score.
The test: Ask what would have to be true in ninety days for the engagement to have been wrong.
The four questions that decide the shape
Three or four yeses means you need an orchestration rather than a tool. Most buys in this category should be orchestrations and end up tools, because nobody asked these in order. The longer treatment is here.
- 1
One task, or a domain?
Drafting an email is a task and a tool solves it. Local presence across seventy-five locations is a domain, and a tool will handle a slice of it while the coordination problem stays exactly where it was.
- 2
Do the outputs need to share context?
If the review-response system does not know what the master record says, its outputs will contradict your listings. Shared context is the difference between several tools and one system, and it is an architectural commitment rather than a feature.
- 3
Is brand voice or compliance a real constraint?
Regulated verticals and brand-at-scale operators both need a gate that scores output before it publishes. If neither applies, you need less than you are about to be sold.
- 4
Will this run for more than six months?
Systems that run for quarters accumulate organisational memory — approved and rejected outputs, drift corrections, tuned thresholds — that compounds. Systems that run for weeks do not, and should be bought as tools.
A three-week sequence
Weeks 1-2 — Inventory and disagreement count
Catalogue every source system and measure how much they disagree. This is unglamorous and it determines everything downstream.
Week 2 — Shape decision
Four questions, answered against your actual stack rather than a maturity model. The output is a decision, and it is allowed to be "you need a tool, not us."
Week 3 — Ranked sequence and governance design
The build order with dependency reasons, and the thresholds and routing tiers that will govern it. Both, at once, before anything is built.
Then stop
A strategy engagement should end with a decision, not roll into a build on the same invoice. If the same firm builds it, that should be a separate decision made with the document in hand.
The last one matters more than it reads. An engagement that rolls straight from strategy into build on the same invoice has a structural reason to recommend a build.
When you should not buy one
- You already ran a serious internal assessment and have a ranked list you believe. Pay for execution instead.
- The bottleneck is approval latency, not production capacity. No AI strategy fixes a four-day legal review; that is an operating change and it is free.
- You have one location, one system, and one person doing marketing. Buy tools, use them, revisit in a year.
- Leadership does not agree on what the company sells or to whom. That has to be resolved by the founders first, and no external strategy can substitute for it.
Common questions
- What should an AI strategy consulting engagement deliver?
- Five things. A ranked build sequence with the dependency reason attached to each position, not a list of opportunities — a strategy that could be executed in any order is a list. A source-system inventory covering every system holding a customer record, location record, or piece of brand truth, including the spreadsheets nobody admits to. A shape decision: tool, orchestration, or operating change, since these are different purchases with different failure modes. The governance design — thresholds, routing tiers, what gets logged — produced alongside the build rather than after it. And one measurable number with a baseline and a date to be judged against.
- How long should an AI strategy engagement take?
- Roughly three weeks for most mid-market and multi-location operations. Weeks one and two on the source-system inventory and the disagreement count, which is unglamorous and determines everything downstream. Week two on the shape decision, answered against the actual stack rather than a maturity model. Week three on the ranked sequence and the governance design together. Then it should stop and end with a decision, rather than rolling into a build on the same invoice — if the same firm builds it, that should be a separate decision made with the document in hand.
- How do you know whether you need an AI tool or an AI orchestration?
- Four questions. Is this one task or a domain — drafting an email is a task a tool solves, local presence across seventy-five locations is a domain where a tool handles a slice and leaves the coordination problem intact. Do the outputs need to share context, since a system that does not read the master record will contradict your listings. Is brand voice or compliance a real constraint requiring a gate that scores output before publish. And will this run beyond six months, because systems that run for quarters accumulate organisational memory that compounds while systems that run for weeks should be bought as tools. Three or four yeses means an orchestration.
- When do you not need AI strategy consulting?
- When you have already run a serious internal assessment and have a ranked list you believe — pay for execution instead. When the bottleneck is approval latency rather than production capacity, since no strategy fixes a four-day legal review and that is an operating change you can make for free. When you have one location, one system, and one person doing marketing — buy tools and revisit in a year. And when leadership does not agree on what the company sells and to whom, which has to be resolved by the founders before any external strategy can help.
- How do you evaluate an AI strategy consultant?
- Ask what goes first in their sequence and why the next thing cannot go first — that separates an order from a list. Ask them to argue for the cheaper shape; a consultant who cannot has one product. Ask what the auto-publish threshold will be and in what units, because vagueness there means governance is a later phase and therefore a later invoice. Ask what would have to be true in ninety days for the engagement to have been wrong. And ask for an engagement where the honest recommendation was to do less.
The version we run
Three weeks, three deliverable checkpoints, one decision at the end — including the recommendation not to proceed when that is the honest answer. It ends with the document rather than rolling into a build.
Ready to talk instead? Book the 30-minute consultation. Thirty minutes, no deck.
Related: the governance half · when you need an operator instead · the three-question diagnostic