Patterns, described by the problem they solve.

Any B2B seller

"Are we named when a buyer asks an AI?"

A fixed set of buyer questions, run across ChatGPT, Claude, Gemini and Perplexity, repeatedly, on a date you can point to. You get the appearance rate, the sample size, the citation list, and a re-run later against the identical set — so the comparison means something.

Any B2B seller

Never retrieved, or retrieved and passed over

Two failures that look identical from outside. One is a placement problem, the other is a page problem, and they need opposite work. Guessing which you have is how a budget gets spent on the wrong fix for three months.

Agency / B2B sales

Sending infrastructure that reports its own health

Domains, SPF, DKIM and DMARC, warmup, pool rotation, bounce and complaint monitoring. The point isn't that it sends — it's that you learn it's degrading before your reply rate tells you.

Agency / B2B sales

Dedup and suppression, before the spend

A hard gate checking every lead against everyone you have ever contacted — before you pay for the record and before it enters a sequence. Re-hitting the same inbox is how sender reputation quietly dies.

Any

"We're out of leads" — usually untrue

Before buying more data, prove the shortage exists: sample the reserve, push it at the platform, count how many are accepted as genuinely new. Inventory is a claim. Loadability is a measurement.

Property / lettings

Landlord follow-up that remembers the last call

Context assembled from the CRM before dialling; objections, promises and documents requested stored per contact; outcome written back to the system of record; conflicts flagged for a human instead of silently overwritten.

Regulated industries

A deterministic guard on what the agent may say

Prices, fees, percentages and contractual commitments checked against a client-approved allow-list in code — not by a second model. A probabilistic writer needs a deterministic gate; a reviewer that agrees confidently is worse than none.

Any outbound team

A voicemail is not a qualified lead

An agent that logs an answering machine as interest will quietly manufacture work for your sales team. Detection, classification and an honest outcome schema belong in the build, not in a later fix.

Finance / compliance

A filings-and-disclosure agent with a paper trail

In a regulated workflow the answer is only half the requirement. Every retrieval, every source document and every model decision is logged so a reviewer can reconstruct how a conclusion was reached — and the agent refuses rather than improvises when the source isn't there.

Agency / consultancy

Per-client AI spend, and who owes what

Two problems that get confused. Your product’s own model calls are tagged at the point of the call, so every request already knows whose it was. Your developers’ AI tools — Claude Code, Cursor, Copilot, loose API keys — are attributed to the client project afterwards, out of band, with nothing sitting in a developer’s way.

Where a figure is inferred rather than measured it says so, because a number you cannot defend in a billing conversation is worse than no number. Fixed platform overhead stays separate from client-driven work — billing the first as the second is how trust goes. The output is a breakdown you can bill from.

Any team running agents

A budget that refuses, not one that alerts

For calls your own product makes, an alert only tells you the money is already gone. A gate in front of the request estimates the cost, reserves it, and declines the call if it would breach the cap. The difference shows up the night an agent gets stuck in a loop.

This belongs in your product’s path and nowhere else. Nobody should put a proxy in front of an engineer’s editor. Ceilings on developer tools work through key limits and approvals, not by intercepting requests.