Ideas and lessons from building accountable AI
Practical guidance on contracts, AI governance and better decisions, from the team that built CIP.
Latest insights
Why AI needs a human sign-off
AI can read faster than any team. It cannot be accountable. How to keep a named person in charge of every decision.
Read article →
The gap between approving a decision and committing to it
A payment, a price change or a contract obligation is approved on one day and executed later. Conditions can change in between.
Read article →
AI agents are starting to change your systems. Who approves?
AI is moving from giving recommendations to taking actions. Approval has to move with it.
Read article →
Approve first, then act: governance before access and change
Examples from healthcare, finance and government of a simple rule: nothing executes until the right person has approved it.
Read article →
Why transformation stalls: scale platforms, not applications
New technology often arrives faster than the organisation can run it. Strong foundations fix that.
Read article →
Customers judge a platform when something fails
Features win the demo. Reliability, recovery and readiness win trust in production.
Read article →
Five risks to check in every supplier contract
Most contract surprises come from the same few clauses. A short checklist for commercial and procurement teams.
Read article →
Audit-ready decisions: keep the evidence as you go
Rebuilding a decision trail months later is slow and unreliable. Capture it while the work happens.
Read article →Why AI needs a human sign-off
Elias Kareem
AI tools now draft, summarise and flag issues in seconds. But when a finding is wrong, a regulator, a customer or a board will not ask the software what happened. They will ask who approved it.
Rules are moving the same way. The EU AI Act asks for human oversight of high-risk AI systems, so people can understand, check and override what the system produces. Some dates in the Act may still change, so check current guidance for your case.
What a good sign-off looks like
- A named reviewer. Every result has an owner, not a team inbox.
- Evidence beside the answer. The reviewer sees the source text, not only the AI summary.
- Approve, change or reject. The person can correct the result, and the change is recorded.
- Nothing leaves before approval. Reports go out only after sign-off.
This is how CIP works: AI finds the points, a person checks and approves each one, and the record shows who decided and when.
The gap between approving a decision and committing to it
Elias Kareem · First shared on LinkedIn
Many business decisions, such as financial obligations, pricing commitments and actions based on documents, go straight from approval into the systems that execute them. Finance and operations systems then carry them out.
The risk sits in the time between the two. A decision approved last week may no longer fit this week's budget, policy or prices.
Add a checkpoint before commitment
- Re-check at the point of execution. Compare the approved decision with current conditions before it is committed.
- Stop or return, don't guess. If something has changed, send it back to the approver.
- Keep both records. What was approved, and what was finally committed.
Our business applications add this checkpoint: payments before settlement, pricing decisions before they go live, and document-based decisions before action.
AI agents are starting to change your systems. Who approves?
Elias Kareem · First shared on LinkedIn
Until recently, AI mostly suggested. A person read the suggestion and acted. Now AI agents in everyday tools such as Microsoft 365 can start workflow updates themselves, across finance, commercial and operations work.
That creates a new split between starting a decision and committing it to a system. Tools that manage agent identity and permissions are a good start, but they do not decide whether a particular action is right for the business today.
Four questions before an agent goes live
- What may it do alone? Keep that list short.
- What needs approval? Anything that commits money, changes a contract or reaches a customer.
- Who owns it? Each agent needs a business owner who answers for its results.
- Can you replay what happened? Keep a record of inputs, steps, approvals and outcomes.
Approve first, then act: governance before access and change
Elias Kareem · First shared on LinkedIn
When a request comes from a person or an AI agent, the safest design puts a preparation step between the request and the system that carries it out. Separation of duties, approval paths and audit records are applied in that step.
Three examples
- Healthcare. A request for access to a clinical system is prepared, but access is granted only after Information Governance approves it.
- Finance. A pricing change or new obligation is checked against company policy before it is committed.
- Government. An action linked to an official document goes ahead only after the delegated approver has signed it off.
The work still moves quickly. The difference is that execution happens only when approval exists, in the same context and with a record of who gave it. This is what our Decision Orchestration application does inside Microsoft 365.
Why transformation stalls: scale platforms, not applications
Elias Kareem · First shared on LinkedIn
AI assistants, automation tools, analytics and digital channels are often introduced one at a time. Each works on its own, but together they create duplicated tools, mixed experiences and growing complexity.
Organisations that scale well take a different path. Instead of adding applications, they build a platform that new capabilities can plug into. They focus on:
- AI capabilities that can be reused
- Data that is governed from the start
- Integration through clear interfaces
- An operating model that can grow
- A consistent experience for users
- Clear ownership and accountability
The result is not only faster change. It is the ability to keep adding new capabilities without complexity growing at the same rate. The winners in AI will not have the most tools. They will have the strongest foundations.
Customers judge a platform when something fails
Elias Kareem · First shared on LinkedIn
Companies invest heavily in features, AI and user experience. These matter. But customers rarely judge a platform while everything is working. They judge it when something goes wrong.
A solution can look excellent in a demonstration. The real test comes when customers depend on it every day.
Three questions every platform should answer
- Can problems be found quickly?
- Can failures be recovered from safely?
- Can the support team respond with confidence?
As AI and digital investment grows, these abilities matter as much as the technology itself. Building software shows technical skill. Building a platform that customers can rely on shows maturity. This view comes from our own work building and running CIP.
Five risks to check in every supplier contract
Elias Kareem
Supplier contracts are long, but problems tend to sit in the same places. Before you sign or renew, check these five areas.
- Renewal and notice. Does it renew on its own? How much notice must you give, and by when?
- Liability and indemnities. Is there a cap, and does it match the value and risk of the work?
- Termination. Can you leave if service is poor? What does exit cost, and what happens to your data?
- Data and confidentiality. Where is your data held, who can see it, and what must the supplier do after a breach?
- Price and payment. Can prices rise during the term? Are payment terms and service credits clear?
Write down what you found, who reviewed it and what you decided. That record helps at renewal and in any dispute.
This checklist supports commercial review. It is not legal advice.
Audit-ready decisions: keep the evidence as you go
Elias Kareem
When auditors, boards or customers ask why a decision was made, many teams search through email, shared drives and chat. The answer is often incomplete.
It is easier to capture evidence at the moment of decision:
- What was reviewed, with the exact document version.
- What was found, linked to the source text.
- Who decided, what they chose and when.
- What happened next, so outcomes can be measured.
CIP keeps this record for each contract it reviews, so the trail is ready when someone asks.
How we built CIP
Elias Kareem
CIP shows how we design, build and run an AI platform. Here is what we set out to do, what we built and what our quality reviews show.
Challenge
Reviewing long business documents takes skilled time, and results can vary between reviewers. A summary alone is not enough for important work: people need the evidence, a clear review step and a record of the decision.
Solution
We built a secure cloud platform, with each customer's data kept separate, that moves every document through a governed workflow.
A clear workflow
Each job moves through stages from submission, classification and analysis to human review, decision, report and notification.
AI with evidence
AI reads the document. AI and rule-based checks find risks, facts and recommendations, with quotes from the source.
People decide
A reviewer approves or rejects the analysis, and the decision is recorded. Users have separate roles.
Built to run
Services scale down when idle and wake when work arrives. Reports are produced as PDF, Word, PowerPoint and JSON.
Outcome
In our own tests on a fictional two-page IT services contract, checked against an answer key with eight expected risks, CIP found 8 of 8 on 2 October and 7 of 8 on 15 October 2026. These are our own tests, not customer results.
What we learned
Test more than one run
Results varied between the two tests on the same contract, so we measure consistency, not just a single good result.
Check the evidence
We test whether important risks were found, quotes match the source and severity is right.
Keep improving
We have since made further changes, including work on consistency, and are testing them.