Before connecting AI to your business, make sure it can find the right information, understand your working rules and distinguish current decisions from outdated records.
Last checked: 13 October 2026
Before connecting AI to your business, make sure it can find the right information, understand your working rules and tell current decisions from outdated records. You probably do not need a new platform for this. You need to know which records are authoritative, who owns them and what AI may access.
Imagine asking an AI assistant to prepare a client report.
The approved template is in SharePoint. The client history is in a CRM. Current work is recorded in a project management system. A colleague updated the reporting procedure, but an older version remains in a shared folder.
Which information should the AI trust? A colleague who knows the business might recognise the outdated document. An AI assistant needs that context made available. Otherwise, it can produce a polished report based on the wrong instructions.
This is not a theoretical obstacle. In a government-commissioned survey of 3,500 UK private-sector businesses with at least five employees, 21% identified data complexity, including insufficient or unstructured data, as a barrier to AI adoption. The interviews took place in 2025, and the findings are self-reported (Department for Science, Innovation and Technology (DSIT), AI Adoption Research, published 2026).
At WeGetDesign, we approach AI adoption from the information and process side first. We build and manage software, and we have connected AI tools to our own project management platform. Organising business knowledge is an operational decision before it becomes a technology decision.
Most businesses already store information somewhere: Microsoft SharePoint, Google Drive, Notion, shared folders or department-specific systems. The question is whether anyone can reliably identify which information is correct.
Consider three versions of a customer onboarding policy. One is approved, one contains useful working notes, and one describes a procedure that has been replaced. Giving AI access to all three does not resolve the disagreement. It gives the assistant three possible answers.
A useful structure needs basic rules:
This does not mean buying a new knowledge-management platform. Existing systems may be adequate once their structure and maintenance rules improve.
Documents describe how work should happen. Operational systems show what is happening now.
A project management system may hold task owners, progress updates, decisions, risks and outstanding actions. For an operations director, those records may be more immediately useful than another static handbook.
Suppose you ask an AI assistant which projects need attention this week, and why. A handbook alone cannot answer that. The assistant needs current project records and agreed criteria for identifying a concern. It must also communicate uncertainty: an overdue task might signal a delivery risk, or its date might simply be wrong.
Business knowledge is therefore a combination of documents, structured operational records and working rules.
WGD Projects is our project and delivery management platform. It holds structured records about projects, tasks, updates, documents, risks and delivery activity.
We have connected WGD Projects to Claude (Anthropic's AI assistant) across 12 organisation boards, with each board exposing 31 defined operations: 16 that read records and 15 that create or change them. We counted these on 13 October 2026 from the connector's own tool listing. They describe what is available, not how much is used or what it saves.
Here is one use we can show. On 13 October 2026, working on the board for our own internal playbooks project (no client data), Claude turned a plan agreed in conversation into eight tickets, each with an opening update written to our house convention. As work landed, it moved the first ticket to Done and the second to In progress. The plan was agreed by a person before any ticket was written.
The connection has safeguards we rely on. Updates are internal by default; making one visible to a client needs an explicit choice. Changes can be previewed before they are written, and repeated requests are not applied twice.
Our task-writing convention matters as much as the connection. Every update begins with a plain-English explanation, followed by technical detail where needed. That makes records easier for colleagues to read and gives the AI a consistent structure.
The principle transfers. A consultancy might apply it to delivery updates, a property business to maintenance work, or a membership organisation to outstanding applications. The software matters less than the structure, ownership and access rules.
You do not need to reorganise the whole company before testing AI. Start with one recurring activity, such as a client report, task review or internal policy question.
Before you start: Name the process owner and a human reviewer. Identify the records you are authorised to use. Do not introduce confidential or personal information into a new AI service without assessing and approving its use.
Map the documents, customer records, task boards, templates and other systems the activity depends on.
Done when: every essential input has a known location and owner.
Find conflicting versions, abandoned folders and informal instructions. Ask the responsible person to resolve disagreements. Do not delegate company policy decisions to AI.
Done when: the team knows which records take precedence and how uncertainty is escalated.
Agree simple conventions for document names, approval status, effective dates, task descriptions and completion criteria. Markdown can be useful for readable instructions, but format alone does not make information accurate.
Use this: our free knowledge-base starter gives you a record template, a table of which record is authoritative for each topic, instructions for the assistant and the ten-question sheet from step 5. Setting it up for one process takes about 45 minutes if you already know where its documents live.
Done when: someone unfamiliar with the process can find the current instructions without asking several colleagues.
Review the data involved, the service being considered and the permissions required. A colleague's ability to read a file does not by itself settle whether that file may be shared with an external AI provider. Start with the minimum access needed and distinguish reading from changing records.
If a connection gives AI access to everything a person can see, then treat that person's access as the AI's access, and narrow it before connecting.
Done when: the proposed information access has been reviewed and approved.
Prepare ten questions whose answers can be checked against approved records. Include straightforward cases, conflicting documents, missing information and questions that should be referred to a person. Record the source used, whether the answer was correct and what needed correction.
Human checking is already common: in DSIT's 2025 survey, 67% of AI-using businesses reported significant human checking of outputs. That figure describes respondents' practices, not evidence that their checks were sufficient (DSIT, AI Adoption Research, 2026).
Done when: a named reviewer has recorded all ten results and the changes needed before broader use.
Policies change, tasks close and responsibilities move. Decide who maintains the authoritative records, how corrections are made and when they will be reviewed.
Done when: someone is accountable for keeping the information current.
Once information is dependable, a business can introduce reusable instructions, sometimes called skills or playbooks, for recurring activities.
A client-reporting playbook might specify which records to consult, which template to use, what information must be excluded and who approves the final document. It keeps the process consistent while each report still reflects the individual client.
At WGD, documented working conventions sit alongside our connected systems. A playbook can guide the assistant, but it does not replace enforced permissions or human accountability.
If business knowledge exists only inside one assistant's conversations, changing providers becomes harder. Keeping authoritative records in systems the business controls gives the organisation more options.
Different AI tools may be able to use the same underlying records through suitable connections. That does not make models interchangeable. Each model and integration must still be tested for answer quality, access controls, performance and cost.
The objective is to compare tools against the same business task, not to rebuild the business around each new assistant.
An impressive AI demonstration does not automatically mean an improved business process.
In DSIT's research, 75% of AI-using businesses reported improved workforce productivity, while 77% reported no change in revenue. These are self-reported perceptions, not independently measured gains or proof that productivity caused revenue outcomes (DSIT, AI Adoption Research, 2026).
For a useful evaluation, compare the process before and after AI assistance. Record the time spent, including checking and corrections; the accuracy and completeness of the result; the number of revisions; staff approval requirements; and operating costs.
These measures will not necessarily move together. A report may be faster to draft but slower to verify. A project summary may be quick to produce but miss an important dependency. Measure the work that matters before claiming savings.
Organised records can still contain mistakes. Policies can be ambiguous. Permissions can be misconfigured. AI outputs can vary between models or between runs.
Changes made through the connection are recorded under the person who connected it, so we review them as our own.
Some activities require specialist oversight, particularly when personal data, regulated decisions or contractual obligations are involved. Clear information structures reduce confusion; they do not remove the need for security, testing or professional judgement.
We have not measured time or cost savings from our own connection, and this guide claims none.
Before buying another AI platform or connecting every company system, choose one activity your team understands.
Find the information it uses. Identify who owns it. Make the records consistent. Decide what AI may access. Test ten questions against known answers. Keep a person responsible for the result.
That first exercise will tell you more about your organisation's readiness than a polished demonstration using generic information. If you want a second pair of eyes on it, we are happy to talk it through.
Our team can help you navigate these trade-offs and build a governed solution.
Whether you’re replacing spreadsheets, modernising internal tools, or planning a complex platform, our process protects you from building the wrong thing.
No obligation. No pressure. Just clarity.