Roadmap Only or Build Team: The Fractional Hire Question

Buy strategy on its own only if two things are already true. You have engineers with genuinely free capacity next quarter, and there is a named person who can sign the AI budget without convening a committee. If both hold, a roadmap is exactly what you are missing and paying for delivery you already own is waste.

If either one is false, a roadmap-only engagement does not give you a plan. It gives you a queue. The queue sits behind the roadmap in a shared drive and the roadmap ages badly, because a use-case list written against last quarter's model prices and last quarter's tooling is stale in about ninety days.

Most mid-market companies fail the test. That is not a character flaw, it is arithmetic: a fifty to three-hundred person company usually has an engineering team already committed to the product that pays the bills, and no slack to spend on an internal AI project that has never had an owner.

Both start the same way. Two to six weeks of discovery, interviews across the departments that think they have an AI problem, an audit of what data exists and who owns it, and a scored list of candidate use cases. Any competent operator produces roughly the same artifacts here, which is why the two shapes are so easy to confuse when you are reading proposals side by side.

The fork is the week after the roadmap is presented.

Roadmap only hands over the document and moves to a lighter cadence: a monthly review, vendor calls, governance edits, someone to argue with when a department head wants to buy something silly. The work is real and the person is often excellent. They just do not build.

Roadmap plus build team treats the document as an input rather than the deliverable. The same firm takes use case one and puts it into production, with the leader who wrote the roadmap accountable for whether the thing they recommended actually works.

The tell is in the statement of work, and it is one word: deliverables. If every deliverable is a document, a session, a framework or a review, you are buying shape one no matter how the proposal is titled. Ask what runs at the end and who is on the hook when it breaks at 2am. The interview version of that test goes through what a good answer and a bad answer sound like.

It dies quietly, and the published numbers on this are worse every year rather than better.

S&P Global Market Intelligence surveyed more than a thousand organizations across North America and Europe and found the average one had scrapped 46% of its AI proofs of concept before they reached production. In the same study, the share of companies abandoning most of their AI initiatives jumped to 42%, up from 17% the year before. That is not a slow decline. That is a market discovering, all at once, that the hard part was never the idea.

BCG's survey of a thousand CxOs across 59 countries landed in the same place from a different angle: 74% of companies had yet to show tangible value from AI, 26% had built the capabilities to move past proofs of concept, and 4% were consistently generating significant value across functions. MIT's NANDA study put it most bluntly, reporting that 95% of enterprise organizations had gotten zero return despite US companies spending somewhere between $35 and $40 billion.

None of those failures were caused by a bad roadmap. They were caused by nobody building what the roadmap said.

Here is the finding that should decide this for most readers, and it is uncomfortable for anyone whose entire offer is a document.

MIT's NANDA researchers looked at what separated the AI projects that reached production from the ones that did not. Buying from a vendor who customized the system, or partnering with one, was roughly twice as successful as building internally. In the same work, only 5% of custom enterprise AI tools survived the trip from pilot to production. The study drew on 52 structured interviews, a survey of 153 business professionals and analysis of more than 300 public AI deployments.

Read those two findings together and the roadmap-only pitch gets awkward. It leaves you on the internal-build path by default, which the best available evidence says fails about twice as often, and it leaves before the consequences arrive.

Accenture's CEO Julie Sweet described the same pattern from the other end of the market, telling analysts the firm had "lots of clients who have started things on their own and then come to us who've got good proof of concept that their team was able to do but then just can't scale it". Those companies did not lack a plan. They lacked the second half of one.

BCG's rule for how AI leaders allocate resources is the single most useful number in this argument: 10% into algorithms, 20% into technology and data, and 70% into people and processes.

A roadmap-only engagement sells you a slice of the 10%, plus a bit of opinion about the 20%. The 70% is workflow redesign, integration into systems that were not built for this, evaluation, the retraining of people whose jobs change, and the unglamorous work of getting one team to use one thing every day. That is where AI programs get won, and it is the part a document cannot do for you.

Which is why the price comparison people make between the two shapes is the wrong comparison. The real one is the roadmap-only fee plus everything you still have to buy afterward, against the bundled fee. Run it that way and the gap usually inverts.

Real assets, and one expensive problem.

The assets are genuine. A scored use-case list with the pet projects already killed. A vendor shortlist that saves you three months of demos. A written policy on what staff may paste into a public model. A business case for use case one that finance can read. If you have never had an owner for any of this, a good fractional AI officer working in advisory mode is worth the fee for the arguments they settle alone.

The problem is the sentence at the end: now build it. Deloitte surveyed 3,235 leaders across 24 countries and found insufficient worker skills to be the biggest barrier to integrating AI into existing workflows, which is the polite way of saying most companies do not have the people. Hiring them is not quick and it is not cheap: Indeed's data puts the average US base salary for a machine learning engineer at $188,730, before employer costs, before equity, and before the months of search.

So the honest accounting for shape one is the fee, plus a hire or a build vendor, plus the calendar time to find either. If your plan for the roadmap is "we will find someone," you have bought a document and an intention.

Systems that run, and a set of downsides worth naming out loud before you sign anything.

On the good side, by day ninety you have use case one in production, an evaluation suite that tells you when quality drifts, a runbook for the failure modes, and a leader whose recommendations have been tested against reality rather than defended in a meeting. Here is that quarter broken out week by week, including the gate at week six that says whether the rest of it is worth paying for. That last part matters more than it sounds. An advisor who has never had to make their own advice work in your systems is guessing with confidence.

On the bad side, three things. You now depend on a firm that knows how your integration works better than you do, which is fine until you want to leave, so make source ownership and handover explicit in the contract rather than pleasant in conversation. Second, scope drifts toward whatever the firm is good at, because every shop has a favorite hammer. Third, the person grading the work is the person who did the work.

That third one is the real objection to this model, and it deserves its own section rather than a bullet.

A firm that sells you the roadmap and then sells you the build has an obvious reason to write a roadmap full of builds.

We are that kind of firm, so treat this as a disclosure rather than neutral commentary. It is also the most predictable way this arrangement goes wrong, and the pages selling fractional AI leadership almost never mention it, which tells you something about who those pages are written for.

Four controls make the incentive manageable, and any decent operator will agree to all four without flinching:

1. Price the roadmap separately, and pay for it separately. A discovery phase bundled into a delivery contract is not a decision point, it is an onboarding step.
2. Put a real kill gate at the end of discovery. You should be able to take the roadmap and walk, with no penalty and no awkwardness, and the contract should say so in words.
3. Require the roadmap to name at least one buy-instead-of-build option per use case. If nothing in a fifteen-item roadmap can be solved with software that already exists, the roadmap is a sales document.
4. Ask what they have talked a client out of. An operator who has never killed a use case has never had the conversation that earns the fee.

Gartner's count of who is actually selling what gives that last point some teeth. Of the thousands of vendors claiming agentic AI capability, Gartner reckons only about 130 offer genuinely agentic features, and it expects more than 40% of agentic AI projects to be canceled by the end of 2027 on cost, unclear value or weak risk controls. Somebody has to be willing to say no to a use case, and if it is not the person you hired, it will end up being your CFO twelve months later.

Answer these about the quarter you are actually in, not the quarter you keep promising yourself.

Do you have engineers with genuinely free capacity next quarter? Free means not on the product roadmap, not on the migration, not on call. Twenty percent of somebody's Friday is not capacity, it is a hobby.

Has that team put an AI system into production in the last year? Not a prototype, not a demo for the leadership offsite. Something with error handling, evaluations and a person whose phone rings when it breaks. If the answer is no, the first build will take longer than anyone in the room thinks, and it belongs to the group MIT found reaching production about half as often as vendor partnerships did.

Is there a named owner who can sign the AI budget? One person, with a number, who can say yes on a Tuesday. If the answer involves a steering committee, neither shape will help you yet, and buying either one is an expensive way to discover that.

Yes to all three, buy the thinking and keep the building. No to any of them, buy the shape that comes with people. And if the third question is the one that fails, fix ownership before you spend anything, because a roadmap handed to nobody is the most expensive PDF your company will ever commission.

The two shapes are usually presented as a choice. They work better in sequence.

Buy the discovery on its own, at its own price, and buy it from someone who can build. Then take one use case off the top of the roadmap and have them ship it while the discovery is still warm. You get the honest version of both shapes: an independent-ish plan, because you could still walk after the gate, and proof that the person who wrote it can make their own advice work.

What that looks like on our price list, since a post arguing about money should show its own: a paid look at what is actually automatable before anyone writes a roadmap, a scoped starter build at $1,500 to $2,500 fixed, and a full workflow live in production in about two weeks at $5,000 fixed. Retainers start at $5,000 a month for companies that want ongoing build-and-run, and plenty of clients never need one.

Rates for part-time leadership move on scope and seniority rather than on a published rate card, so the numbers behind a monthly retainer are a separate argument worth reading before you compare two quotes.

The build half is normally an engineer embedded in your team rather than a project handed over a wall, because the integration knowledge is the expensive part and it should end up inside your company.

Whoever builds it, insist that an evaluation suite ships with the system. A model that quietly degrades and nobody notices is worse than no model, and evals are the cheapest insurance in this entire category.

Sometimes the correct move is to buy nothing this quarter, and any operator worth hiring will tell you so on the first call.

Skip both if there is no named owner. Skip both if your data lives in one person's spreadsheets and nobody has agreed which system is the record. Skip both if the reason this is on the agenda is that your board asked what your AI strategy is, because the answer to that question is a decision, not a purchase.

And skip both if you are under twenty people. At that size a fractional executive is a layer of management for a company that does not have management, and one well-scoped automation will teach you more than any roadmap. Buy the build, skip the strategy, and revisit the question when the second department starts asking. The rest of the cases where buying leadership is the wrong move carry prices, which is the part nobody selling it publishes.

Nobody selling either shape has an incentive to say that, which is why you rarely read it. McKinsey's survey found 88% of organizations regularly using AI in at least one business function, while nearly two thirds have not begun scaling it across the enterprise and only 39% can point to any EBIT impact from it. A share of that gap is companies who bought leadership before they had anything for a leader to lead.