The short answer
Consultancies deliver recommendations, agencies deliver software built to a brief, freelancers deliver small scopes cheaply, and in-house delivers alignment at salary cost and hiring risk. They differ most on three axes buyers under-weight: what physically exists at the end, who owns it, and whether the engagement is designed to finish. The comparison usually omits a fifth option, embedded fixed-scope engineering, which is the one built around finishing.
Four ways to get AI built, four different incentive structures, and a buyer usually comparing them on day rates, which is the one axis that matters least. Here is the comparison on the axes that decide outcomes.
The comparison
| Consultancy | Agency | Freelancer | In-house | |
|---|---|---|---|---|
| Deliverable | Analysis, strategy, recommendations | Software built to your brief | Software built to your spec | Whatever your roadmap says |
| Cost structure | Time and headcount, open-ended | Per project or per sprint | Day rate or per project | Salary plus ramp time |
| Timeline to production | Production is out of scope | Months, brief-dependent | Weeks to months, scope-dependent | Quarters including hiring |
| Ownership at the end | You own a document | Contract-dependent; hosting often theirs | Usually yours, rarely documented | Yours by definition |
| Fails when | Nobody builds the plan | The brief was wrong | Scope exceeds one person | Hiring misses or work runs out |
The consultancy, honestly
Consultancies sell judgment, and good judgment is real value: market context, risk framing, political cover for a decision already half-made. The structural problem is that the engagement completes at the recommendation, which is exactly where the documented failures begin. Somebody still has to build the thing, integrate it, and carry it into production, and by then the consultancy is in the next engagement. Buy consulting when what you lack is a decision. Do not buy it when what you lack is a system.
The agency, honestly
Agencies build, which already beats slideware. The weakness is positional: they build from briefs, at a distance from your workflow, and misunderstood problems are the first root cause of AI project failure in the research. The brief is a guess written by whoever attended the kickoff. Watch also for the retention design: hosting in their accounts, prompts undocumented, a system only they can operate. If you use an agency, force the six-part handover into the contract before the build.
The freelancer, honestly
The right senior freelancer is the best value in the market for a small, well-specified scope. The limits are structural: one person means no bench when scope grows, and production disciplines, security review, evals, runbooks, monitoring, are exactly what a solo builder under time pressure rationally skips. Freelancers shine at prototypes and well-bounded tools; treat production hardening as a separate, deliberate step whoever builds the first version.
In-house, honestly
Alignment is perfect and everything else is expensive. Deployment-capable AI engineers are the market's scarcest tier, being absorbed by labs and hyperscalers at compensation operating businesses rarely match, and a first hire learns your domain on your payroll. In-house is the right end state for companies where AI work is continuous. It is a slow and risky way to get a first system built.
Day rates are the axis that matters least. What exists at the end, and who owns it, decide the outcome.
The fifth option the comparison omits
Embedded fixed-scope engineering: senior engineers inside your workflow, a production system in weeks, and a complete transfer at the end. It borrows the consultancy's judgment (discovery happens in your operation, and sometimes concludes "do not build"), the agency's delivery, and none of the open-endedness: the engagement is paid to finish. It is our lane, so weigh the framing accordingly, then hold it to the same tests as everyone else: the nine questions apply to us too. And if the roles rather than the firms are what confuse, the role-by-role comparison untangles those. And if you are shopping this market from London specifically, the local tier map saves a day of tab-opening.