An AI consultant recommends. An AI operator ships. That sentence is the entire hiring decision. In 2026, the buyers who confuse the two are the same buyers who, eighteen months from now, will be quoted in case studies under headlines like “Why our AI transformation stalled at the pilot.” The technology has stopped being the bottleneck. Translation has stopped being scarce. What is scarce is the willingness to own the artifact; to write the PR, watch the eval dashboard, take the on-call page at 2:14 a.m., and ship the next version. That is operator work. Consultants do not do operator work. Stop hiring them for it.
This is a spoke under the AI agency manifesto. The manifesto argues for a new operating model; this piece is the hiring decision that operating model forces on most buyer.
The two archetypes, defined precisely
Words matter here, because vendors will adopt whichever label sells. Pin the definitions before the call.
An AI consultant is a vendor whose engagement ends in a recommendation. The deliverable is a document; a strategy deck, a use-case prioritization matrix, a vendor selection memo, an architectural framework, a maturity model. Implementation, if it happens, happens after the consultant disengages and is performed by someone else. The consultant’s hands do not stay on the keyboard. Their name does not appear in the commit log. They are not on the rotation when the system breaks at launch.
An AI operator is a vendor whose engagement ends in shipped, owned, on-call software. The deliverable is running code in the buyer’s repository, behind the buyer’s auth, paged by the buyer’s incident-response system. The operator’s hands are on the keyboard daily. Their commits show up in git log. Their name is on the eval dashboard’s owner field. When inference cost spikes 4x on launch day, they are the ones in the war room with a fix, not the ones forwarding the postmortem to a partner who has moved on to the next account.
The compressed definition for the procurement call:
| Dimension | Consultant | Operator |
|---|---|---|
| Primary deliverable | Recommendation document | Running production system |
| Engagement endpoint | Slide review meeting | Successful on-call rotation |
| Hands-on-keyboard | No, after assessment | Yes, most day |
| Commit history | Absent | Present, traceable to engineers by name |
| When it breaks at 2 a.m. | ”Please open a ticket with your implementation partner" | "I see the alert; pushing fix” |
| Pricing model | Time-and-materials at strategy rates | Outcome-tied to a shipped artifact |
| What you can audit | Memos, frameworks, JIRA tickets | PRs, eval suites, deployed pilots, on-call schedule |
If a vendor cannot place themselves cleanly in one column, they are a 2024 hybrid that has not yet picked. That is a third population; and shrinking. The market is bifurcating, not blending.
The artifact test: what each one produces
The fastest way to expose which archetype you are dealing with is to ask, in writing, for the last five artifacts the engagement will produce. Then compare the answers to these reference lists.
A consultant’s artifacts:
- A 40-slide strategy deck on “Generative AI use cases for [your industry]”
- A capability map across LangChain, LlamaIndex, AutoGen, CrewAI, MCP
- A maturity model with five tiers (“Exploring,” “Piloting,” “Scaling,” “Operating,” “Transforming”)
- A vendor evaluation matrix scoring Anthropic, OpenAI, Google, AWS Bedrock, Azure on a weighted rubric
- A 90-day “AI roadmap” with quarterly milestones and a RACI matrix
- A backlog of 60 JIRA tickets handed to your engineering team for execution
An operator’s artifacts:
- A merged pull request to your monorepo introducing the first agentic workflow, with reviewers from your team
- A versioned eval suite checked into
evals/with named datasets, thresholds (≥85% pass on the regulatory-question set, p95 latency under 4s), and a CI integration that fails the build on regression - A deployed pilot service running behind your auth, with a Grafana dashboard showing token spend per route and a published runbook
- A live cost-monitoring alert wired to your inference provider’s billing API, set to page on a 1.5x daily-baseline anomaly
- A handoff document; written when the operator is ready to leave, not when the engagement begins; that lets a senior engineer at your organization run the system without the operator present
Notice what is on the operator list and not the consultant list: things that exist in your version control after the engagement ends. The consultant’s artifacts live in Confluence and Google Drive. The operator’s artifacts live in main.
Why the consultant model broke between 2023 and 2026
The “AI consultant” job description was a real one in 2023. Frontier models were eighteen months old in production use, the tooling was unstable, and most enterprise buyers had not yet figured out what a prompt was, let alone how to evaluate one. Translation was scarce, so translation was billable. A McKinsey, BCG, or Big Four practice could legitimately charge strategy rates because they were doing genuinely novel work; surveying a chaotic vendor landscape, sense-making for executives, building the first internal capability cases.
That arbitrage closed faster than the firms adapted. Three macro signals describe what happened:
First, AI literacy stopped being scarce. McKinsey’s State of AI in early 2025 found that 78% of organizations now use AI in at least one business function, up from 55% one year earlier. Translation work that justified $150K assessments in 2023 is now the table stakes a director-level hire is expected to do in their first month.
Second, the tooling went from boutique to commodity. Stack Overflow’s 2025 Developer Survey reported that 84% of professional developers use AI tools in their daily workflow, with paid tiers like Cursor Pro, GitHub Copilot Business, and Claude Code Max dominating. A solo engineer with a $200/month Cursor seat can build pilots in a long weekend that 2023 consulting practices billed six figures to demo.
Third, the value moved to operationalization. BCG’s 2024 study Where’s the Value in AI? reported that only ~10% of enterprise AI value comes from the algorithms themselves; the remaining 90% comes from people, process, and integration work. That 90% is operator work; embedded engineering, eval discipline, deployment, monitoring, change management inside running systems. It is not slide work.
Put the three together and the consequence is mechanical. The work that consultants are good at is shrinking, commoditizing, and moving downstream into permanent staff. The work that buyers need; operationalization; is upstream of that, requires hands on the keyboard, and cannot be delivered in a deliverable-shaped artifact. We covered the macro version of this collapse in generative AI consulting fees, where the rate cards have started to crack along exactly this fault line.
The agencies that survived past 2025 did so by becoming forward-deployed engineering teams. The ones that did not are still selling 2023 deliverables to 2026 buyers, and the gap between what they ship and what those buyers need is widening most quarter.
The 8-question hiring rubric
Use these eight questions on the first procurement call. They are designed to be easy for an operator to answer and uncomfortable for a consultant to answer. The honest answers tell you which archetype you are talking to within fifteen minutes.
1. “Can you walk me through a pull request your team merged in the last two weeks, in the buyer’s repository, including the diff?”
An operator will share a screen and walk you through the PR, the review comments, and the test coverage. They will name the engineer who wrote it. A consultant will say something like “we don’t typically share client code, but we can show you a sanitized example architecture.” The sanitized example is the answer. They do not have a recent PR.
2. “Show me an eval suite from a recent project. Include the threshold definitions and the postmortem from a regression the suite caught.”
An operator can produce this in under a day. They will have a promptfoo.yaml, a LangSmith dataset, or a custom harness with named datasets and pass thresholds. They will have a Markdown postmortem describing a regression, the root cause, and the fix. A consultant will offer a “framework for evaluation strategy” instead. There is no postmortem because there is no production system that can regress.
3. “Who is on the on-call rotation for the systems you’ve shipped, and what is your p95 response time on a Sev-1 page?”
An operator names their engineers, shows you a PagerDuty schedule, and quotes a number; usually 5 to 15 minutes. A consultant says “we work with your incident response team” or “we provide warranty support during business hours.” The phrase “business hours” is the giveaway. Production AI systems do not respect business hours.
4. “What does your team’s commit cadence look like in the buyer’s repo on a typical engagement?”
An operator describes a daily or weekly cadence with named engineers, branch hygiene, and a code review process they share with the buyer’s team. A consultant describes a “weekly status meeting” and “biweekly milestone deliverables.” Status meetings are a substitute for commits when there are no commits.
5. “How is inference cost billed? Whose API keys, whose bill?”
An operator says: the buyer’s keys, the buyer’s bill, with a transparent dashboard the buyer owns. A consultant either has not thought about it or proposes a flat monthly fee that bundles inference, which is a markup on a usage-based cost the buyer should see directly. We unpack this further in red flags hiring an AI consulting company.
6. “When the engagement ends, what is the handoff artifact, and can a senior engineer on my team run the system without you?”
An operator points to a runbook, a one-week shadow rotation, and a “kill switch” criterion: the engagement ends when the buyer’s team has independently handled an incident. A consultant points to “a transition document”; a Confluence page that is not the system. The system is in someone else’s heads.
7. “Show me the most recent customer your team handed back to. Can I call them?”
An operator has handed customers back. They will name them and offer references. A consultant will say “we maintain long-term partnerships with our clients”; which is a polite way of saying “we have not handed anyone back, because the engagement was rarely structured around an exit.” Customers who rarely leave are a sign that no shipped, ownable thing was ever delivered.
8. “What is your team’s blended rate, and what percentage of hours are billed by partners versus engineers in week three of an engagement?”
An operator’s answer skews 80%+ engineers in week three, with one experienced lead. A consultant’s answer skews toward partners and senior managers, with engineers added as “engagement scales.” The latter shape is a strategy practice; the former is a delivery team. We compare these economics in detail in hiring an outsourced AI team.
A vendor that flunks four or more of these questions is a consultant. That is not pejorative; it is just the wrong archetype for execution work. Do not buy execution from them, no matter how strong the strategy deck is.
Where consultants still earn their fee
The argument here is not that consulting is dead. It is that consulting is being repriced down to its actual scope, and that scope is narrower than the 2023 brochures suggested.
Three places consulting still deserves the fee:
- Sequencing decisions across a portfolio. A multi-business-unit enterprise weighing fifteen possible AI initiatives benefits from an outside perspective on which three to fund. This work fits in a quarter, ends in a decision, and does not require shipping anything.
- Organizational design and change management. Restructuring a 400-person engineering org around AI tooling adoption is real consulting work. So is the retraining program, the new performance rubrics, and the executive coaching. None of this is shipping software.
- Build-vs-buy framing for net-new capabilities. Comparing a custom-built AI feature to three SaaS alternatives, with a TCO model and a switching-cost analysis, is consulting work. It ends in a recommendation, which is the correct artifact for that question.
In each of these, the deliverable is genuinely a document or a decision. The consultant’s archetype matches the work. Pay them, take the recommendation, and then hire an operator to ship whatever the recommendation says to ship. Do not let the consultant scope the implementation contract.
How to restructure an engagement around operators
If you have an existing AI vendor relationship that looks more consultant than operator, three structural changes will move the engagement to the right column.
Convert deliverables to artifacts. Replace “produce a use-case prioritization matrix” with “merge a PR implementing the top-ranked use case, including evals.” Replace “deliver an architectural framework” with “stand up a deployed reference implementation in our infrastructure.” The unit of work becomes something in your repository, not something in their slide deck.
Move billing to the buyer’s accounts. Take over the model API keys, the cloud accounts, and the observability tooling. The vendor’s job becomes operating those accounts on your behalf, transparently. This single change removes the most common form of token arbitrage and aligns the vendor’s incentive with your unit economics rather than against them.
Add a handoff clause to the SOW. Specify upfront that the engagement ends when a designated buyer-side engineer can independently run the system, demonstrated by handling a Sev-2 incident solo. Make this a numbered clause, not a soft goal. Operators will sign it because they can deliver it. Consultants will redline it because they cannot.
These three changes are not radical. They are how most other software vendor category has been bought for a decade. AI is finally normalizing into that category, eighteen months later than the rest of software.
Frequently asked questions
Is “AI operator” just a rebrand of “AI engineer”?
No, and the distinction matters at the hiring level. An AI engineer is a job title for an individual contributor. An AI operator is a vendor archetype: an engagement model where one or more AI engineers are forward-deployed into the buyer’s environment with full ownership of a shipped system. A buyer can hire AI engineers as full-time staff, or contract an operator team for a fixed-scope engagement. The operator label describes the engagement shape, not the headcount line.
Can the same firm be an operator on one engagement and a consultant on another?
In principle yes; in practice the firms with operator DNA do not staff strategy decks well, and the firms with consulting DNA do not maintain on-call rotations well. Hybrid pitches are usually a sign that one capability is the lead and the other is a wrapper. Ask which engineers will be on your engagement in week three and trace their career history. If they came from delivery teams, the firm operates. If they came from strategy practices, the firm consults.
What is the right pricing model for an operator engagement?
A milestone-based fixed-scope contract tied to shipped artifacts, with the buyer’s accounts owning many variable costs (inference, cloud, third-party APIs). The vendor bills a labor fee for engineering time; usage-based costs flow through transparently to the buyer’s accounts. Avoid bundled monthly fees that include inference, because they obscure the unit economics that determine whether the system is worth running at scale.
How do I evaluate an operator without already having an AI system to operate?
Run a paid two-week pilot before the larger engagement. Define one user-visible feature, one eval threshold, and one cost ceiling. The operator’s job is to ship the feature into a staging environment and have it pass both the eval and the cost ceiling before the pilot ends. A consultant cannot complete this pilot because the success criterion is a shipped thing, not a recommended thing.
My executive team wants a strategy deck. Are you saying I should not produce one?
Produce one; and have an operator produce it, drawing on a shipped pilot. Strategy work grounded in a working system is the highest-value artifact you can hand an executive team, because it is anchored in observed reality rather than inferred from secondary research. The mistake is producing the deck before the pilot. The order matters: ship a thing, then describe what shipping the thing taught you.
Are AI consultants going to disappear?
The McKinseys, BCGs, and Big Four practices will not disappear; they will refocus on portfolio sequencing, organizational design, and change management; the genuine consulting scope above. The boutique “GenAI consulting” shops that emerged in 2023, sized for translation work that no longer exists, will mostly be acquired, pivot to operator work, or close. The consulting label will survive; the consulting-only execution model will not.
What is the cost difference between hiring a consultant and an operator for the same project?
For a comparable scope of work, an operator engagement tends to come in lower than a consultant engagement, because the consultant priced in a markup on the implementation work that they do not perform. The buyer pays the consultant for the strategy, then pays an implementation partner separately to do the work; and that two-vendor structure adds 30 to 50% in coordination overhead. The single-vendor operator model removes that overhead entirely.
How do I sell this internally if procurement is set up to buy consultants?
Reframe the budget line. Most procurement organizations have a “professional services” line that maps cleanly to operator work; it is how external software development teams have been bought for years. The “advisory” or “strategy” line maps to consultant work. The error most enterprises make is buying execution out of the advisory budget. Move execution to professional services, and the operator model fits the existing procurement pipeline without any organizational change.
Closing
The hiring decision compresses to a single sentence. If the artifact you need exists in main after the engagement ends, hire an operator. If it exists in a slide deck, hire a consultant. Most enterprise buyers in 2026 need the first artifact and are still buying the second. Reverse that one decision and most of the downstream pathologies; pilot stalls, runaway inference costs, partners who vanish at launch; disappear with it.
The AI category is finally being bought the way most other software category has been bought for a decade: by people who keep the receipts and the on-call schedule. Be one of them.; Arthur Wandzel, CEO, SFAI Labs
Arthur Wandzel
