Start with useful work
Agentic AI can carry a process forward across several steps: gather evidence, retain context, prepare the next action and recognise when someone needs to decide. Making that useful in a business begins with a clear output and an accountable owner.
For a first deployment, choose a bounded workflow. “Help with sales” is difficult to assess. “Prepare an opportunity assessment from approved sources, with evidence and unresolved questions” gives a reviewer something concrete.
Define what completion means before connecting the agent to business systems. A useful assessment might contain dated sources, the reasons an opportunity fits, delivery concerns and a recommendation to proceed, investigate further or pass.
Bid research in practice
At AI Venture X, our internal bid workflow is one example. A specialist AI agent researches and qualifies opportunities, develops the case for a bid and maintains context as work progresses. [1]
Research and preparation
- Investigate the opportunity and buyer.
- Assess relevance and identify uncertainties.
- Prepare evidence for a decision.
- Keep track of the next steps.
Human decisions
- Authorise buyer contact.
- Approve prices and commitments.
- Decide whether to pursue the bid.
- Approve the final submission.
The operating principle is bounded autonomy: give the agent room to prepare useful work while keeping consequential commercial decisions under human control. This describes our internal approach; it is not a claim of independently measured client savings.
Separate preparation from permission
A proposed email, a price recommendation and a submitted bid carry different consequences. The ability to prepare an action should not automatically confer permission to execute it.
OWASP identifies excessive functionality, permissions and autonomy as sources of excessive agency. Its guidance includes limiting tool access, enforcing permissions in downstream systems and requiring human approval for high-impact actions. [2]
A practical approval boundary
Review the actual message, recipient and commercial terms. If those change after approval, check whether the revised action needs a new decision. Keep a record of who approved what and what was executed.
These controls need to exist in the workflow and connected systems. A prompt asking an agent to be careful cannot substitute for access control.
Test what happens when things go wrong
A successful demonstration is only the beginning. Before a workflow gains more authority, test cases such as:
- A source is incomplete or contradicts another source.
- The opportunity changes after a recommendation is prepared.
- A reviewer withdraws approval.
- A tool times out after an action may already have succeeded.
- The workflow reaches a processing or spending limit.
Specify the expected response: stop, ask, retry safely or reconcile the result. A timeout must not become an automatic instruction to repeat a potentially completed action.
An inspectable engineering reference
Our public Python reference explores exact-action approvals, expiry, revocation, tenant boundaries and recovery after a lost response. It uses a simulated agent and mock CRM. The reference demonstrates those local controls, not production performance or model accuracy. [3]
Measure useful output and review effort
Before a pilot, record how the current process performs. Agree the measures and unacceptable errors with the person who owns the work.
- Preparation time: how long it takes to produce a reviewable assessment.
- Evidence quality: whether material claims have accessible, relevant sources.
- Review effort: the time people spend checking and correcting the output.
- Recovery: whether uncertain or failed actions are resolved without duplication.
Compare similar work and include rejected or abandoned cases. Faster drafting can be valuable, but it does not by itself establish better commercial decisions.
From a workflow to a working system
Our approach to agentic AI consulting starts with the business problem, the people responsible and the evidence needed to judge the result. We then assess the options, design the workflow, connect it to the right systems and support adoption. [4]
For UK organisations exploring enterprise AI automation, a useful first discussion is specific: what should the system prepare, what may it execute, what must someone approve and what would justify extending its scope?
Discuss a workflow with AI Venture X
Bring one business process, the outcome you need and the decisions you want to keep under human control.
Book a 30-minute scoping call ↗Explore our AI systems and servicesSources and further reading
- Richard Russell: AI staff in practice. Founder’s account of internal bid research and human approval boundaries.
- OWASP, LLM06:2025 Excessive Agency. Guidance on capabilities, permissions and autonomy.
- Enterprise Agentic AI: public Python reference. Runnable examples and documented simulation limits.
- AI Venture X: our delivery approach. Discover, assess, design, connect and embed.