Getting more from the systems you already run

Most businesses are not ready for AI. Their processes are not defined or structured enough for it to earn its place: the steps change from person to person, the rules live in people's heads, and nobody has written down what a good outcome looks like. Enterprise pilots still fail at high rates, and the cause is usually the foundation underneath the model rather than the model itself.

The good news is that the value rarely needs a new platform. It sits in the systems a business already pays for: the shared inbox, the CRM, SharePoint, WhatsApp and the spreadsheets that grew up around them. AI reads, sorts, answers and routes the work people currently handle by hand, writes the results back into those systems, and passes anything that needs judgment to a person. The skill is in finding the right place to start, and waiting is sometimes the right call.

Here are four use cases from four different businesses. The systems and the problems differ, but each followed the same three steps. First, understand the problem as it really runs. Then scope the smallest build that proves the value. Then deliver it into the tools people already use.

Documents that file themselves

Understand. Invoices, claims, onboarding packs, contracts, delivery notes and tax documents arrive as password-protected PDFs, phone photos and scans bundled several to a file. In most companies a skilled employee still opens each one, works out what it is and whose it is, and retypes the contents into a system of record. Watching that person work is the fastest way to find the real specification, because the rules they apply in their head are the rules the system will need.

Scope. Anything that is really policy is decided up front rather than left to the model: a document that contains another document is still the document it is. The list of types lives as data the business owns, so adding one is an administrator task rather than a rebuild. And a document the system cannot place is never filed on a guess. It becomes a visible "needs review" item, with its best guess attached, for a person. We also choose the model on evidence: candidates are graded against hand-checked examples, and in one test a mid-tier model matched a top-tier one on accuracy at 2.3 times lower cost.

Deliver. For one client, we built a document extraction pipeline that has now processed more than 100,000 of their documents. They arrive by email, chat, client portals and watched folders. Each is unlocked if it is protected, read, classified against the client's own list of types, and has its fields extracted into a structured record in the system the business already uses. Accuracy and cost were measured, not assumed: in one extraction run for that client, covering tens of thousands of records, accuracy checked by hand came to 99.3%, and the compute for the entire run cost about $216.

The extracted values keep records current, drive reporting and flag the cases that need a conversation. The team no longer opens every attachment. They work from the short list the system could not place. Because these pipelines handle sensitive records, we design data security into the architecture from the first day.

A CRM that keeps itself fresh

Understand. Client records go stale for a boring reason: updating them is a chore. A new proof of address arrives, someone files it, and the address on the record stays the old one. Six months later nobody trusts the data, and the CRM the business paid for stops being the source of truth.

Scope. For an advisory business, we put an assistant in a panel beside the client record. We began strict, with a handful of document types: the ones handled most often rather than the most complicated, because the value comes from volume. Each type defines the fields to read, the folder it files to and the name it is given. Nothing is saved until a person confirms it, which keeps a named human accountable for what is filed. Identifying documents that arrive with no client attached was scoped as the next slice, with its own size.

Deliver. A document is dropped into the panel. The assistant works out what it is, reads the fields that type defines, and shows them next to what the record currently holds. A person checks, and only then is the file stored and the record updated.

Every document is categorised against the business's own catalogue, with the reasoning shown, so a person is confirming a judgement rather than accepting a verdict. It is stored under the same name, tags and folder as every other document of its type, so anyone can find what someone else filed. The confirmed values are written back to the CRM, and where one disagrees with what is already held, both are shown side by side and nothing is overwritten quietly. Because the catalogue is data, a new document type is added by an administrator with no development work.

Customer conversations on WhatsApp

Understand. For many businesses, WhatsApp is where customers already are. It is also where repetitive questions, after-hours messages and promising leads pile up, leaving companies to choose between hiring more staff and letting enquiries wait.

Scope. An intelligent WhatsApp agent draws its answers only from the company's own approved material and can take only the actions on an agreed safelist. Every conversation, document and qualified lead is recorded in the existing CRM, such as Microsoft Dynamics 365. Questions that need expertise go to the right person with the full conversation attached, and the same pattern extends to Microsoft Teams and the web.

Deliver. Our launches typically begin with a soft pilot and staff on standby, so the business sees real conversations before it relies on the system. Over two months in production at a financial services firm, one of our agents handled 8,188 real cases, resolved 90.6% without a human agent and returned 1,154 hours of staff time to the team, about 577 hours a month, at roughly $0.09 per case.

Spreadsheets and licences, in one place

Understand. In a recent Blueprint for a large UK property company, individual staff had signed up for their own AI tools to get through the day, which helped them personally and gave the business no shared standard or platform. The company was paying roughly £40,000 a year in licences for fairly basic analytical work, and senior team members spent 147 hours a month on high-volume manual tasks across spreadsheets with no single source of truth.

Scope. The build replaces the licensed tools and pulls the work into one place, with direct connections to the specialist platforms the company already relies on, so data flows without being re-keyed.

Deliver. The rebuild replaces the licensed tools for about £2,000 a year in hosting. Manual effort is projected to fall to 18 to 23 hours a month, human review included, which returns 124 to 129 hours every month.

What the four have in common

Different systems, different teams, one method. Each began with a process that already had clear rules, high volume and a cost the business could measure, because the return depends on choosing the right problem. Each was scoped small enough to ship quickly and big enough to matter. Each wrote its results back into a system the business already ran, and each kept a person in charge of anything uncertain.

They also passed the same readiness test. If the person doing the work can explain the rules, a machine can follow them. If they cannot, it is not an AI problem yet but a process problem, and the honest answer is to fix that first.

That method is the AI Blueprint

Understand, scope, deliver is how our AI Blueprint works. It starts with a four-week Readiness Sprint that maps how your processes actually run, identifies the highest-value opportunities and picks the one worth building first. The build phase that follows puts production AI workflows directly into the tools you already use, at a fixed price.

You leave with a working prototype of your highest-priority use case, human-in-the-loop approvals with a full audit trail, and a training and rollout plan. It suits organisations that want working AI in their workflow without waiting months for separate discovery and build phases. And if the honest answer is that you are not ready yet, we will tell you that too.

The Claude Partner Network

Everything described here is about to reach further. On 2 October 2026 we were certified in Anthropic's Claude Partner Network, which extends what we can offer to company-wide Claude rollouts, new agents for WhatsApp, Teams and the web, and Claude embedded directly in systems such as Dataverse, CRMs and service desks. For our clients, each of the use cases above can now extend across more of the business.