A plan to reach AI Readiness for an AI-native consultancy —
from an AI Engineering perspective.

01 · Executive Summary

The firm sells itself as an AI-native consultancy, but underneath the brand the engineering practice is inconsistent.

Every project rebuilds the same building blocks, and the best of them are buried inside individual solutions where no one can see, review, or reuse them.

The only AI Engineering solution in production for customers is a low-code, high-level agent — not the deeper engineering capability that would put the firm on the leading edge of the market and deliver on its stated mission: guiding human organizations through AI-driven change, from roadmaps to impact.

The gap between what the firm sells and what it can reliably build and maintain is the single biggest risk to the firm's 2026 plans.

02 · Strategic Directions & Plan

The whole plan in one view — five directions to commit to together, and the phased roadmap that delivers them. More than tools or process, the aim is a shared shift in mentality and culture — becoming an AI-native company across the board, where every team builds and works with AI as a default way of working.

The plan — build the foundation first; the rest are outputs

The plan front-loads one real investment: Step 1 — establishing how the firm builds AI, and the technical capacity to do it. Everything after it — the first proof, reuse, a packaged offering, and scale — is an output of that foundation, not separate work. This is what directly addresses the challenges in §3: no reuse, the credibility gap, talent risk, and the maintenance gap.

03 · Current State

1. Wasted effort & no reuse ("mono" agents)

Every AI project rebuilds the same common building blocks from scratch:

• The result is several implementations of the same capability where there should be one best version — and the team never collaborates to build the single, top-tier solution for each block.
• The best work stays buried inside individual solutions, invisible at the top level, so no one can see, review, or improve it.
• Engineering collaboration suffers: feedback, code reviews and shared technical expertise are hard to build when nothing is reused or visible.

2. Low-code is a real business line carrying layered risk

Low-code agents are a genuine opportunity and should continue — but today they are the firm's only production proof to customers, and they carry four stacked risks:

3. Internal tools aren't built to compound

• Internal tools (the flagship internal agent, the internal copilot experiments) are siloed, with no shared AI strategy.
• No framework absorbs common challenges once, so the next build is no faster — effort doesn't compound.
• No dogfooding discipline: the firm doesn't prove its own agents beat off-the-shelf tools.

4. The credibility gap

• The firm markets deep AI capability while the engineering foundation is inconsistent — the brand promise and the delivery reality have diverged. This is the central commercial risk.
• Gartner expects most early agentic implementations to miss their targets on integration, governance and talent gaps — not model quality Gartner (forward-looking).
• The firms that win are the ones with a real platform underneath.
• The consequence: not building something better than commercial tools for the firm's own work undermines its credibility and its literacy when selling AI services.

5. The organisation can't make time

• By the nature of consulting, engineers are pulled across client work — it's hard to carve out time to build a shared stack, platform and technical foundations.
• Without a shared mentality, common stack and engineering practices, there's no foundation to build on — so it's easier to keep starting from scratch, and the gap compounds.
• Every other challenge is downstream of this — a platform that is "everyone's job in their spare time" never ships.
• The fix (§9): a protected platform team that treats the platform as a product — the precondition for everything else.

04 · Strategic Vision & the AI Ecosystem

The goal isn't "an agent builder." It's a complete AI ecosystem — where the platform, the way of building, and a shared mentality stack into something that unblocks the challenges by nature, not by heroics.

What the ecosystem unblocks — by nature

05 · Proposed Framework & Reference Architecture

A modular agent platform is built from small, reusable pieces assembled like a puzzle — and it stands on a handful of shared pillars, each a capability the whole team builds on instead of rebuilding per project.

06 · Recommended Common Stack & Tooling

The firm delivers through two complementary paths — meeting customers where they are, while building its own depth and IP. Two ways to deliver, one engineering practice.

The in-house platform — a suggested reference architecture

Quality & evals — the one place not to cut corners

Evals are automated quality checks for AI — known-good examples with a clear pass/fail, run on every change so nothing ships that quietly got worse. It's also critical to monitor cost and latency, not just quality.

• Measure quality, don't assume it. AI fails silently — confident but wrong — so dashboards look normal while the model is wrong Pan.
• LLMs are black boxes. Observability (tracing) is the only way to see what an agent actually did, and to debug it Braintrust.
• Feedback interfaces feed monitoring automatically. In-product ratings, thumbs and corrections flow straight into observability — turning real usage into new test cases and feeding the cognitive loop (§7) Langfuse.
• Start small, from real failures. A handful of real failure cases beats a giant test suite; grade the outcome, not a rigid step-by-step path Anthropic.
• Let AI help grade — carefully. An AI can judge open-ended answers, but only once it has been checked to agree with a human expert Husain.

07 · The Cognitive Loop / Consultant Digital Twin

The platform unlocks this stage. Once it exists, the firm can close the cognitive loop — turning its human capital (workflows, domain knowledge, continuous-improvement (Lean) discipline, client context, taste) into "token capital" that compounds. This is where the firm gains its real edge, starting from how it operates internally: a consultant digital twin built by the consultants themselves — an evolving, dynamic system. A platform is replicable; this encoded judgement is not.

How the loop is operated, concretely

• Encode: every consultant's interaction and feedback builds the agents — so each can deploy their own 1:1 digital twin, carrying their workflows, judgement and domain know-how.
• Run: those blocks power both internal delivery and client agents.
• Feed back: real traces, corrections and failures flow back into the prompts and evals — the twin sharpens.
• Govern it: the loop only compounds if evals are broad and the firm dogfoods the same thing customers get.

08 · Service Offering & Sales Enablement

Sales talk about AI but don't know what to actually offer. Rather than inventing a fixed catalog up front, the firm lets the offering emerge: as it builds the platform and dogfoods it, the capabilities that prove genuinely valuable become what it sells — discovered through delivery, not guessed.

Early signals point toward low-code data agents, bespoke production agents, a Day-2 maintenance service, and sovereign / AI-Act-ready architecture — but these are directions to validate, not a menu to commit to now.

It also compounds: each customer implementation unblocks a cluster of adjacent use cases the firm can replicate and offer with little extra effort — every engagement widens both the catalog and the sellable offering.

Cost efficiency, not cost reduction

"Cost reduction" reads as headcount cuts and clashes with people-first brand values. Cost efficiency — the same people delivering more, better — is on-brand, aligns with continuous-improvement (Lean/Kaizen) principles, and is what the platform actually enables. This framing should be the standard sales line.

Dogfooding as the sales engine

If the firm can't build something its own teams prefer over off-the-shelf ChatGPT / Claude / Gemini, it cannot credibly sell AI services.

• Forces the platform to be genuinely good.
• Teaches sales the real capabilities from experience, not slideware.
• The flagship internal agent — make it beat off-the-shelf tools for the firm's own work, and sales sell from lived proof.

Enablement — a practical path from awareness to mastery

Skip passive awareness sessions. Follow a practical path — awareness → understanding → mastery, where people learn AI by building with it, not by sitting through slides:

• Hands-on labs — learn on the platform, not on theory.
• Internal hackathons — feed winning patterns straight into the catalog.
• Solution champions / ambassadors / evangelists — drive peer adoption; Applied Delivery Engineers anchor it.
• Coaching over decks — enough real, hands-on training to turn people into regular users.

09 · Operating Model: How the Roles Support Each Other

The Forward Deployed Engineer profile — why and how

An FDE (Forward Deployed Engineer) is, simply, an engineer who embeds directly with a customer to build and ship their solution on the ground — the model OpenAI and Anthropic now use to deliver AI on customer infrastructure.

For such a firm, this lower-tech FDE owns low-code, hands-on client delivery and the customer relationship — unblocking delivery capacity while keeping its core AI engineers on frontier work. A clear career ladder, not a low-code dead-end:

A probable addition — Prompt Engineer

A likely future role: a Prompt Engineer who can spin up quick POCs and handle day-to-day maintenance — tuning prompts, evals and context. Lower barrier to enter, high-leverage for fast validation and upkeep, and it frees both FDEs and core engineers from routine work.

10 · EU Compliance & AI Sovereignty as a Differentiator

Build the platform sovereign by design — able to run independently of the providers that feed it — and the firm gets two things at once: it eliminates the risks recent events have exposed, and it opens commercial doors that vendor-locked competitors can't.

Why — what recent events proved

• In June 2026, the US government ordered Anthropic to cut off all foreign access to its top models (Fable 5 / Mythos 5) — taking them offline worldwide overnight. Even the most advanced AI can be switched off by a single government order Time.
• Under oath, Microsoft France could not guarantee EU data would never be handed to US authorities; the US CLOUD Act compels US providers to disclose data wherever it's stored Register CLOUD Act.
• The ICC was pushed off Microsoft Office onto open-source tools — "digital sovereignty" is now a board-level topic across Europe heise.
• Demand is real: Gartner sees sovereign-cloud spend up ~83% YoY in Europe Gartner.

Sovereign by design — the payoff

Honest caveat: the EU lags on raw AI compute, so the firm differentiates on trust, jurisdiction and agility, not raw capability — don't oversell it McKinsey.

11 · Appendix: References

Claims marked (directional), (provisional), (heuristic) or (forward-looking) rest on single or fast-moving sources — verify before client use. Proposed KPIs (reuse rate, time-to-first-agent) are firm-specific, not industry-standard terms.