Too much has to be remembered
Which rules apply, what was decided last month, who owns which project: the answers are held in memory, and recall is the overhead.
One workspace, any AI tool.
One recurring workflow, run from written rules.
One pipeline, made of steps that compose into the next.
DelfinFlow builds the layer between AI tools and business systems. That layer is the rules, the decisions and the project memory, kept in files the business owns.
Work moves at the speed of whoever has to look it up. One place to update a project and one place to record a decision, whether ten people work on one project or one person works on ten.
Every business runs on a pattern that is rarely written down: how the work starts, who decides, what was settled last time. An agent can take up that work at any moment. Given the pattern as files, what it does is narrowed: the same rules, the decisions in force, the same answer for the same work.
The work arrives as a pipeline of processing steps. Written down once, a step is a step the next pipeline can be built from.
Where the operation leaks
None of these is a competence problem.
Search finds the conversation; the decision in force, what it replaced, and who may change the project have to be reconstructed each time. That reconstruction is the overhead.
Which rules apply, what was decided last month, who owns which project: the answers are held in memory, and recall is the overhead.
Two decisions about the same project disagree, and nothing records which one replaced the other.
A question, its answer and the decision that followed scroll away in a busy channel, and finding them again is work on its own.
What is true for a project sits in a call, a thread or a message, and it is reconstructed again at the next one.
Work that has to move quickly takes the shortest path available, and the shortest path usually leaves nothing written down.
Each tool is capable on its own and works from its own instructions, because none of them is given the same operating model.
Solution
A small operating model for one workflow, kept in the company’s own storage and changed only by review. An assistant reads it and is allowed to do exactly what the model says.
The agent stays general; the workspace is what narrows it to this business’s terms.
The structure follows the published ICM/MWP pattern — context as files, with a person reviewing between stages — and the tooling behind it is built on the JigsawFlow capability contracts.
The workflow written as an operating model.
Business context the assistant answers from.
The assistant that reads the workspace and follows it.
A converged memory, kept current with use.
Nothing changes without approval.
Every decision and change, recorded.
Each step is inspected by the part it has produced, in the way a production line inspects the part it has just made.
Offer
The price covers the analysis of how the business operates and its position against the market, then the design, build and testing of one workflow: the workspace, its memory, the ledger and the assistant that run it.
The base · pilot price
£1,000
one workflow with its memory, on files the business already has
No third-party system integration and no cloud tenancy needed. The files stay where they already are, under the business’s own accounts. Pilot price for the first engagements: the base is a complete unit and is bought without a surface.
A surface, if the workflow needs one+£600 in Teams, +£900 on the company’s own server — or both, £2,500 with the base.
Route A · Microsoft Teams
+£600
£1,000 base + £600 surface = £1,600
the base, delivered where the team already works
Microsoft 365 licensing or pay-as-you-go access is bought directly by the customer; the route required depends on the agent’s data access and is confirmed before setup. One approval from the tenant administrator is required for the read access.
Route B · own server
+£900
£1,000 base + £900 surface = £1,900
the base, served from the company’s own server, with more autonomy
Where there is no other surface to talk to the agent, a simple browser chat is included as a fallback.
Which route: the Teams route is the usual starting point where Microsoft 365 is already in place, because each person reaches the assistant inside the tools already open on the desktop. It is a personal assistant, though, held one to one; the server route is what an agent in a shared channel requires, and the surface chosen follows from whether the workflow is answered per person or per team.Where a team talks somewhere other than Microsoft Teams, another surface is arranged case by case, on the same workspace and the same rules.
Not in this offer: a dashboard, an admin console, a portal, or a custom application.
Engagement
The initial conversation: which workflow consumes the hours, who does it today, what it costs.
How the business operates, and its position against the market. This confirms the scope before the build.
The base alone, or the base with a surface. The base is one week; with a surface, two weeks.
Workflow and rules written with the person who does the job; knowledge base assembled; assistant configured; ledger started. Run first against a copy of the workspace, with the checks green, then on real examples.
A signal to move forward and close the delivery. The workspace speaks for itself through its /help skill: how it runs, how to change it, what it can and cannot do. Feature requests and questions about the operation are answered by the workspace. No dependency on DelfinFlow to keep it alive.
Two weeks, included. Feedback is collected as the team works with the agent; adjustments follow.
After the delivery
Agentic Operations is the business’s own AI infrastructure. It runs on top of the workspace and its integration, adding measurement, monitoring, routing and approval to the system.
The workspace records where runs stall and what needs correcting, then proposes changes to its own instructions — adopted only after the proposed rule passes the cases the delivered workspace already passes. Where a step can be scripted, it is moved into a script or a scheduled job.
Scheduled checks report jobs that did not run, outputs that left the written rules, and decisions that no longer match the work.
An Expert agent holds one role and its scope. A Warden agent holds approval, review and revert. Work outside those roles is dispatched as a single run, carrying the brief written for it — the context is reviewable before the run starts and closed when it ends. Every role and every run is recorded in the same ledger.
Each task is routed by complexity: to the smallest model that can do it, to several models where an answer needs diversity of opinion, or to a deep research methodology that gathers sources before writing.
One browser or native view shows what ran, what was decided, what awaits approval, and the way back.
Connecting the services a business already uses — email, payment gateways, calendars, CRMs and similar — into the operation where they are needed.
Each run is recorded for tokens, cache reuse, cost and time, and compared with earlier runs of the same task. Measured spend is published in the usage benchmark.
Tools, files and data sources are connected through MCP servers, declared once and used by any capable agent. WebMCP applies the same pattern in the browser.
A customer-facing assistant that answers from the business’s knowledge, under the same rules and approval boundaries as the internal operation.
After the delivery · infrastructure
A website, an eshop, a portal, a WordPress installation or the server it runs on is a business asset that changes continuously. Requests for it arrive as a conversation, are prepared as a change, and are applied only after approval. The asset may have been built by DelfinFlow or by another supplier.
Email, chat or a message from the owner becomes a recorded request — what should change, and which part of the asset it concerns.
A Manager process turns the request into a concrete change — page copy, structure, catalogue, prices — and prepares it against a copy of the asset where the platform allows one.
The Warden process shows the prepared difference and applies it on approval. The written rules decide what may be changed, and by whom.
Each change is entered in the ledger with what changed, who approved it, and the path back.
Page adjustments run on the same base workspace as the rest of the operation. The rules, vocabulary and record already written are reused, so the next engagement starts from what exists rather than from a new brief.
Questions
No. A chatbot answers from whatever it happened to retrieve, so what comes back depends on the search. Here the workflow, its rules and its decision boundaries are written down first, and the assistant reads the business’s own guidelines before it starts work and then stays inside them — including stopping when the model does not cover the case. That written model is the deliverable; the assistant is one way to use it.
No. The workspace is not tied to any cloud service or provider: it is files under the business’s own accounts, and the agent runs where the team decides it runs. Microsoft 365 is mostly the chosen AI control surface.
The Teams agent is personal: each person holds it one to one, and it is not a member of a channel, so a team cannot call on it in a shared thread.
An agent that answers in a channel is a separate deployment — a Teams application installed in the tenant, built on the company’s own server or in Copilot Studio. “Route B” provides that deployment, and it is the one that can be given write access to OneDrive.
On the “Route A” agent, no. On “Route B”, only where write access is granted.
A change is prepared as a handoff document and applied by the agent that holds the write permission, so by default nothing is applied automatically. Where an agent is allowed to apply a change itself, the ledger keeps the path back.
In the ledger, with the workspace: every decision and every applied change, dated, attributed and marked with what it replaced. Superseded decisions are kept, never removed, so the state of any project can be read back from the record. A change to files is also a commit in the workspace’s own git repository — a standard history any developer can read.
In the company’s own storage — its OneDrive or its own disk — under its own accounts. DelfinFlow works on it during the engagement and keeps no continuing copy afterwards. Access is read-only wherever the workflow allows it, and nothing is applied to a file without an approved change.
The system belongs to the business and is documented. Two weeks of feedback, review and adjustments follow. After that, further work is a new engagement.
Start
One call, twenty minutes, no cost and no obligation. A deeper analysis follows, written against the workflow as it runs today.