How to recognise an agency that sells slides.

There's a whole industry built on selling tech-lag anxiety. These agencies don't write code. They write 150-page PDF "Strategic Frameworks", wrap them with the word "AI" on the cover, and bill them in six figures. In 2026 the pattern hasn't disappeared: it has learned a new vocabulary. Here is how to recognise it before you sign.

At Jigen we operate in the trenches: we build systems, measure impact on margin, move on to the next client. Yet we're regularly called in second to rescue projects bogged down in classical-consulting mud. In the past two years that mud has changed name — it's now called "AI strategy", "GenAI roadmap", "agentic transformation" — but the structure of the problem is the same: a vendor selling hours of thought instead of systems that run. What follows is a diagnostic checklist built on dozens of projects we've recovered from agencies that hadn't put them into production.

Red flag #1 · The infinite worksite

If the agency's proposal talks about a six, eight or twelve-month "implementation roadmap" to ship version 1.0 of an AI-based system, you're already trapped. The state of the art of models evolves on cycles measured in weeks: an eight-month project is structurally obsolete by month three. You sign on an architecture reasoned around January's models and prices, you go to pilot in July, and best practice has shifted twice.

The reason they sell you eight months isn't technical. It's financial. The longer the timeline, the more hours can be billed for "project manager", "scrum master" and "alignment meetings". A serious agency segments risk: it produces value in isolated three-week units, each with its own end criterion, target number, and closed budget. If they can't deliver an output that moves margin in twenty-one days, they don't know what they're doing. If they know what they're doing but won't commit to three weeks, they're protecting their own revenue. Either way, that's your problem, not theirs.

Quick test: at contract signing, ask the vendor to show you a system they shipped to production in ≤ 21 days for another client. Anonymised is fine. If they don't have the case, they don't have the method.

Red flag #2 · The "proprietary fine-tuning" obsession

The most elegant scam of the decade. You're told your company needs its own exclusive model, trained on your secret data. They promise you an insurmountable competitive advantage.

Training a model from scratch is an exercise for frontier labs. What your company actually needs is a robust retrieval-augmented generation architecture on your documents, paired with the frontier models already available — Claude 4.x, GPT-5, Gemini 3 — orchestrated with well-written prompts. It costs a fraction of the price, takes a tenth of the time, and is infinitely more flexible: when the next model ships, you swap the provider in thirty minutes, not in six months. If the agency pushes fine-tuning before validating a standard retrieval architecture, they're selling useless development hours and locking themselves to a choice that within a year will need redoing.

Real exceptions to the pattern: domains with very tight technical vocabulary (vertical legal, clinical pharma, complex proprietary code) where even frontier models show a measurable gap. There fine-tuning makes sense — but it's a decision taken after proving retrieval isn't enough, not before. If nobody has shown you the numbers of a pre-fine-tuning retrieval system, walk away from the conversation.

Red flag #3 · Zero mention of orchestration

An AI system in production in 2026 isn't "a chat with a model". It's a sequence of specialised agents passing work to each other, calling external APIs, validating data before writing to systems of record, applying routing logic based on task type. If the agency pitch revolves entirely around "we integrate an LLM API into your software", you're buying a wrapper, not infrastructure.

Five technical questions to ask in scoping, in order:

1. Routing across models — which model is called for which task type? (Good example: "long reasoning on Claude Opus 4.7, fast tool-calls on Sonnet 4.6, mass classification on Haiku 4.5 or GPT-5 mini, web search on Gemini 3"). Bad example: "we use GPT-5 for everything".
2. Fallback — when the primary provider degrades or goes offline, what happens? (Good example: "retry with backoff, then switch to a second provider with mapped prompt"). Bad example: "it has never happened".
3. Output validation — who checks that the model's response is well formed, consistent with upstream data, compliant with company policy? (Good example: "JSON schema validated at runtime, fact-check on a second model, refusal rule below confidence threshold"). Bad example: "the model knows what it's doing".
4. Logging and replay — is every interaction logged so the same decision can be replayed a month later? (Good example: "structured log of prompt, model version, context, response; deterministic replay on fixed seed"). Bad example: "we're about to implement that".
5. Cost per interaction — do you know, today, how much a single production call costs you, and how it scales with volume? (Good example: "nine-tenths of a cent per qualified lead, rises to 1.6 cents if the task requires multiple tool-calls"). Bad example: "it depends".

If they stumble on three or more of these five, you know you're facing amateurs armed with handsome slides.

Red flag #4 · "We'll do AI for everything"

A modern variant of marketing-tone is the horizontal promise: "we apply AI to every process in your company". It's the same promise that ten years ago sold "digital transformation": covers everything, commits to nothing, justifies an arbitrary budget. Reality is that AI systems that work are surgical: they take a single hidden-cost process, automate it until the cost disappears, then measure the number. The next one comes after.

Quick test: ask the agency which single process they'd automate first, what number it would move, in how much time, for how much. If the answer is "depends on our discovery", they're selling the discovery, not the automation.

Red flag #5 · Seniors only show up at the pitch

The traditional agency works by inversion: the most experienced seniors sign the proposal, run the pitch, win the trust. Then, at signing, the project passes to junior staff who have never put an agent into production in their lives. The senior will return at the quarterly kickoff to read a status written by someone else.

For SMEs in particular, this is devastating. A task force where the senior architect is also the concrete executor produces qualitatively different output from a chain where the senior is just a commercial ambassador. Quick test: ask for the names of who will write the code. Ask for links to public repositories, technical articles, conference talks. If the agency won't expose them, it's because they don't exist — or they exist but aren't allocated to your project.

Red flag #6 · They deliver slides, not source code

The final deliverable is a PDF. The system runs "on the agency's servers". The code isn't yours. The prompt isn't yours. The database is proprietary. When the agency decides to raise the annual renewal price, you have no alternative: pay or lose everything. It's exactly the same scheme as the SEO agencies of the 2010s that locked clients into proprietary CMSs, rewritten for AI.

A serious collaboration in 2026 always provides three things, in writing in the contract: source code in a repository you own, infrastructure on cloud accounts in your name (even if managed by the vendor), prompts and configurations exportable in standard format. If any of these three is missing, you're building a dependency, not a system.

Red flag #7 · They sell hours, not results

The "time and materials" tariff model dressed as "agile" makes it structurally likely that the project will never end with a system in production. The agency's incentive, in pure T&M, is to continue. The incentive of a vendor that quoted a closed result is to close and move on.

Healthy structures measure on output: a system that does X, in production by Y, fixed price Z, with explicit clauses for who pays scope-creep and who pays technical failures. Structures that measure on input — hours, headcount, sprints allocated — are designed to last.

Red flag #8 · 10:1 ratio between document and code

An empirical metric we use internally when we audit a derailed project: the ratio between pages of documentation (decks, requirements, frameworks, status presentations) and lines of operational code written. Healthy agencies stay below 1:5 — more code than documentation. Slide agencies are often at 10:1 or worse: a hundred pages of documentation for every hundred lines of code. That ratio reads like an ID card: it tells you exactly what kind of work you're paying for.

At signing, ask the vendor for the expected volume of formal documentation (status decks, operational frameworks, stakeholder presentations) versus the expected volume of production code. If they don't have an estimate, they have never thought in those terms.

Thirty-minute test

An operational grid the leadership of a company can run before signing a six-figure quote. Five questions, thirty minutes, no meeting room:

1. Show me a system you shipped to production in ≤ 21 days for another client. Even anonymised. I want to see the architectural schema and the number it moved. Expected time: 5 minutes.
2. Who will write the code on my project, by name and surname? I want links to public repository, talk, technical article. Expected time: 5 minutes.
3. Across the frontier models available today, what routing would you use for my case and why? Expected answer: cost/latency/quality comparison across at least three providers, no monoculture. Expected time: 8 minutes.
4. Is the final deliverable code in my repository, or access to a system you manage? Expected answer: exportable code and configuration, infrastructure on my cloud account. Expected time: 5 minutes.
5. If on day 21 the system isn't running, what happens to the contract? Expected answer: explicit clauses, scope renegotiation, partial refund, clean exit. Expected time: 7 minutes.

If the vendor stumbles on two or more of these five, it's time to close the meeting and look for another vendor. You're about to pay for slides.

If you're already trapped

For those who have already signed and find themselves at five months of "discovery" without a system in production, the option to exit cleanly exists. Three practical steps:

First · written request for code and data. Regardless of what the contract says, a formal request for access to the source code and all data produced so far forces the vendor to make explicit what really exists. It often emerges that the "system" is a series of prototypes never integrated. That's the moment to renegotiate.

Second · set a sixty-day minimum value threshold. If within two months you don't have a system that moves at least one concrete metric — a conversion, an operating cost, a response time — close the contract and cut the loss. Exit clauses exist in all serious contracts; in shaky ones, the threat of public communication is enough to obtain mediation.

Third · operational second opinion, not strategic. Don't call another consulting agency to audit the first. Call a vendor that can ship code to production and have them build, in three weeks, a reduced version of the same system. If they can do it, you know the first agency was scamming you. If even the second can't, the problem wasn't the agency — it was the project's perimeter, and the conversation changes.

Don't be fooled by jargon. Code in production is the only yardstick. If a vendor can't show you operating systems working in the first three weeks of collaboration, stop. The cost of inaction is high, but the cost of chasing empty slides for a year is fatal — and in 2026 competitive advantage is compounded in twenty-one-day cycles, not in three-year plans.

Jigen reading: this checklist doesn't describe an exception in the market — it describes the default. Slide-selling agencies represent today the majority of AI-consulting supply in Europe. Recognising them in thirty minutes is a capability that pays for itself the first time you skip a wrong diligence.