Constraint Engineering: Making AI Diligence Auditable

Last week we described the diagnostic pipeline. This week, the part that makes it usable in a deal process — and the part that still doesn't work.

An unconstrained prompt produces fluent, generic output because the model is drawing from a broad probability distribution across everything it has read. Ask "which rural Altnets are financially stressed?" and you get a competent summary of trade-press coverage. It reads authoritatively. It is not diligence, because you cannot audit a single claim in it.

Constraint engineering is the discipline of narrowing that distribution until the output becomes checkable. Three levers do most of the work:

1. Anchoring every claim to a document

The most effective constraint is the simplest: require the model to cite a specific filing, regulatory publication, or dataset for each assertion, and to state explicitly where data is unavailable rather than estimating. This single rule converts a persuasive narrative into a list of claims a human can verify or reject in an afternoon.

2. Bounding the question to a defined universe

"Rural UK Altnets" is not a universe — it is a vibe. Ofcom's Area 3 classification, active Project Gigabit contracts, and Companies House filing status are universes. Prompts that reference a defined population produce output whose gaps are visible; prompts that don't produce output whose gaps are invisible.

3. Forbidding the inferential leap

Models are strongest at exactly the thing diligence punishes: constructing a plausible causal chain across weakly related facts. An explicit negative constraint — do not infer commercial consequence from a regulatory term without a contractual anchor — removes a whole category of confident, unfounded conclusions.

Here is the honest limitation, and it is the reason we do not describe this as a solved problem.

Constraints reduce variance. They do not eliminate it. Run the same constrained prompt twice against the same dataset and you will not always get the same shortlist back — same facts, different confident conclusions.

That instability is tolerable when the output is treated as a list of things to check. It is disqualifying when treated as a conclusion. The distinction is not academic: an unstable shortlist presented as a finding is how a deal team ends up spending real hours investigating a risk that was never there, or missing one that was.

What actually closes the gap is not a better prompt. It is ground truth — site visits, contract review, conversations with operators and their customers — fed back into the process as evidence that either confirms or kills each generated hypothesis. That feedback loop is the part we are building now. We are not claiming to have finished it.

The prompt architecture decides what gets asked. Evidence from the ground decides what turns out to be true.

One question we genuinely want answered: if you are running AI inside a diligence or research process today, what is your actual sign-off gate — a named human, a score threshold, or something else?

Sources: Ofcom, Telecoms Access Review 2026–31; Companies House; INCA, State of the Altnets 2026. Methodology description is GreySynth's own.

Next: what an AI-native telecom operating model actually looks like in practice.

From Chatbot to Co-Pilot: How Our AI Diagnostic Pipeline Works

Last week we introduced the diagnostic question that should precede every deal thesis. This week: the tool that helps us answer it at scale.

We should be transparent about how this content series was built. The market analysis underlying these articles — the operator profiles, the financial stress-tests, the behavioural-friction patterns — was produced by a diagnostic pipeline that combines generative AI platforms with structured human judgement. The AI is a tool within a human-led advisory process — the same way a financial model is a tool, not the advice. The human judgment, validation, and strategic interpretation are the product. The AI compresses the data-gathering phase.

Here is how the pipeline evolved, and what each stage actually produces:

• V1 — Information retrieval. Broad queries: "list rural UK Altnets and their coverage figures." Output: generic, indistinguishable from published industry reports. Useful as a structured starting point. Not useful as analysis.

• V2 — Contextual synthesis. Geographic filters (Ofcom Area 3 classifications, BDUK Project Gigabit contracts), regulatory overlays, specific audience framing. Output: segmented operator profiles, platform-appropriate formatting. Materially better, but still descriptive rather than diagnostic.

• V3 — Financial diagnostics. Prompts demanding analysis of capital structures: debt-to-equity ratios, interest coverage under high-rate scenarios, ARR vs. build costs. Output: financial stress indicators that highlighted operators with structural funding gaps visible in Companies House filings. The AI's contribution was speed and systematic coverage across dozens of operators simultaneously — compressing weeks of analyst work into hours.

• V4 — Problem diagnostics. The AI shifted from answering "who should we buy?" to surfacing the questions we hadn't thought to ask. This is where the diagnostic-question framework emerged — from the interaction between constrained AI analysis and human strategic judgement, not from either alone.

The lesson is not that AI replaces deal teams. It is that AI, properly constrained, compresses the diagnostic cycle by surfacing patterns in fragmented datasets that human analysts would take weeks to synthesise manually. But "properly constrained" is doing real work in that sentence — the prompt architecture is the product, not the model.

We now maintain a structured prompt registry as core advisory infrastructure, calibrated by deal type and sector.

→ We bring this diagnostic capability to client engagements. Ask us how it applies to your sector.

The Diagnostic Question Before the Deal Thesis

In our first four articles, we mapped the market reality: 19.7 million premises passed, 18% take-up, structural behavioural barriers, and hidden operational frictions that explain the gap. Now we shift from diagnosis to methodology: how should acquirers and advisors think about this differently?

Most fibre M&A strategies start with the wrong question: "How can we acquire the largest physical footprint at the lowest enterprise-value-to-premises-passed ratio?"

It sounds disciplined. It is a formula for buying someone else's customer acquisition problem.

David Schonthal and Loran Nordgren's work on innovation resistance offers a more useful frame. Their core insight: when adoption of a new product or service stalls despite clear rational advantages, the bottleneck is almost never awareness or pricing. It is friction — status quo preference, implementation anxiety, category skepticism — that the provider hasn't identified, let alone addressed.

Applied to fibre M&A, this reframes the entire acquisition thesis. Building on Schonthal and Nordgren's framework, we ask what we call the diagnostic question — the question that must be answered before any tactical plan is generated:

Are we acquiring raw physical lines that are commercially stranded, or are we acquiring a platform with the operational and behavioural capability to overcome the customer inertia holding back rural fibre migration?

This distinction changes every downstream decision. Target selection shifts from coverage maximisation to evidence of customer momentum. Valuation methodology moves from per-premise multiples to take-up trajectory and unit economics by geography. Integration planning prioritises the target's customer acquisition engine — its people, processes, and community relationships — over its physical infrastructure.

The operators who have cracked this tend to share three characteristics: mid-scale regional focus (100,000–400,000 premises), low or zero leverage, and patient institutional backing that allows multi-year customer development without quarterly pressure to show coverage growth.

Competitive advantage does not stem from possessing fibre cables. It belongs to the operator who identifies where the friction points are, reframes the adoption problem, and builds the most effective migration engine for the specific communities it serves.

→ We run diagnostic assessments as part of our deal advisory. Let's discuss your pipeline.

Attribution: Innovation-resistance framework from D. Schonthal & L. Nordgren, The Human Element (Wiley, 2022). Application to fibre M&A is GreySynth's.