The demo worked.
The AI read an invoice, found the supplier, drafted a response, and suggested the next action. Everyone in the room could see the potential.
Then the production questions started.
Which supplier record can it access? What happens when the bank details change? Who approves an exception? Can it post to the ERP? What does it log? How do we stop it? How do we undo a bad action?
That is where the model demonstration ends and the real system begins.
A capable model is useful. It is not a dependable digital coworker by itself. The business outcome comes from the harness around the model: the context it receives, the tools it can use, the permissions it carries, the checks it must pass, the people who approve consequential actions, and the evidence it leaves behind.
If you are moving AI from chat to execution, spend less time asking which model is smartest and more time designing the harness that makes useful work safe and repeatable.
A good answer is not a working process
A chat interface can hide the hard parts of operations.
You type a request. The model produces a polished answer. The interaction looks complete because the visible output is complete.
A business process is different. It crosses systems, people, rules, deadlines, exceptions, and incomplete data. It may start in email, continue in an ERP, require a manager's approval, and finish with an audit record.
Recent reporting on production agent infrastructure describes the required layers as orchestration, memory, context management, evaluations, governance, monitoring, hosting, scaling, and tool integration.[1] That list matters because it exposes the gap between a model response and an operating result.
The model reasons. The harness determines how that reasoning is used.
Without a harness, the AI may know what should happen but cannot reliably prove what did happen. It may produce a recommendation without knowing the latest customer status, call a tool with excessive access, repeat a failed action, or continue when the workflow should stop for a person.
The production question is not simply:
Can the AI do the task?
It is:
Can the business control, observe, recover, and take responsibility for the task when the AI does it?
The nine layers of a production harness
You do not need a large enterprise platform to think clearly about this. Start with nine practical layers.
1. Orchestration
Define the sequence of work. What triggers the agent? Which step comes next? When should it retry, wait, escalate, or stop?
2. Context
Give the agent only the information needed for the current job. Current customer status beats a giant folder of loosely related documents.
3. Memory
Separate durable facts from temporary working state. A customer preference may belong in long-term memory. A one-time calculation probably does not.
4. Tools
Specify which systems the agent may read from or write to. Reading an invoice is not the same authority as changing a vendor record or releasing a payment.
5. Permissions
Use the minimum access required for the role. A digital coworker needs a job description and boundaries, not a master key.
6. Evaluations
Test the workflow against normal cases, edge cases, bad inputs, missing data, and attempted manipulation. Do not evaluate only the model's final wording.
7. Governance
Name the policy owner, the accountable business owner, the approved use, and the actions that require human consent.
8. Monitoring and auditability
Record inputs, tool calls, decisions, approvals, exceptions, and outcomes. If you cannot reconstruct the run, you cannot manage it confidently.
9. Recovery
Design the stop button and rollback route before deployment. A failed workflow should enter a known recovery path, not become a late-night investigation.
This is why harness design is becoming as important as application design. The model is one component. The harness turns that component into a controlled operating capability.
Do not use an agent where a rule will do
Agentic AI is not a replacement for every automation tool.
If a process is stable, deterministic, and fully described by rules, conventional workflow automation or RPA may be the better choice. It is easier to predict, test, and audit.
Agents become useful when the work needs context and judgment across fragmented systems: interpreting an ambiguous request, selecting the next tool, handling variation, or preparing an exception for review. The same production-infrastructure reporting makes this distinction between fixed workflows and cross-system work that depends on reasoning.[1]
The practical pattern is often hybrid:
- Rules handle predictable steps.
- Agents handle interpretation and variation.
- People handle consequential exceptions and approvals.
That is orchestration, not blind automation.
Accounts payable makes the harness visible
Accounts payable is a useful example because it combines documents, ERP data, compliance, fraud risk, and clear approval boundaries.
A recent product announcement described an AP workflow in which invoices arrive through email, portals, APIs, or manual upload. AI agents extract the data, verify vendors and bank details, detect duplicates, validate tax, and perform a three-way match against the purchase order and goods receipt. Items requiring judgment are routed to a person; cleared items are prepared for ERP posting.[2]
Treat that as a workflow pattern, not a recommendation for one vendor.
A sensible AP harness would make each boundary explicit:
- Intake: Accept documents only from approved channels and preserve the original.
- Extraction: Capture invoice fields and attach confidence scores.
- Verification: Compare supplier, bank, tax, purchase-order, and receipt data against trusted systems.
- Exception handling: Route mismatches, new bank details, missing receipts, and unusual amounts to the correct person.
- Approval: Require the named finance approver for consequential decisions.
- ERP handoff: Post only the approved, validated record using narrow permissions.
- Audit trail: Store what was checked, what failed, who approved, and what changed.
- Recovery: Hold or reverse the transaction when a downstream check fails.
The automation handles checks and preparation. People retain accountability for exceptions and payment authority.
That is what a digital coworker looks like in practice: not an AI that does everything, but a role-designed system that completes bounded work and knows when it must stop.
A safer SME pilot
Do not begin with an enterprise-wide AI transformation. Begin with one workflow that has visible pain and measurable outcomes.
Use this pilot structure:
Pick one bounded process. Choose a repeated workflow with enough volume to matter but limited enough to inspect manually.
Map the current path. Document the trigger, systems, decisions, exceptions, approvals, and final record before adding AI.
Assign a business owner. The person who owns the outcome must also own the operating rules. Do not leave accountability with the technology vendor or model.
Create a permission card. List what the agent may read, draft, change, and never do. Name the approval gate for every consequential action.
Build an exception queue. The goal is not zero human involvement. The goal is to move people away from repetitive checks and toward the cases that need judgment.
Test failure and recovery. Use duplicate invoices, changed bank details, missing records, conflicting instructions, and unavailable systems. Confirm that the workflow stops safely.
Measure an operating outcome. Track cycle time, exception rate, rework, error rate, or time spent per case. Do not use the number of AI messages as proof of value.
Review before expanding. Scale only after the workflow produces an audit trail, respects permissions, survives edge cases, and delivers a measurable improvement.
The real upgrade
A newer model may improve reasoning. It will not decide your approval policy, clean your customer data, define your operating roles, integrate your ERP, or take accountability for a payment.
Those are harness decisions.
The companies that move AI into production will not be the ones with the longest list of models. They will be the ones that turn business judgment into context, tools, controls, evidence, and repeatable workflows.
Orchestrate the work. Do not hand the work to a model and hope.
Sources
[1] https://africa.businessinsider.com/local/markets/the-anatomy-of-an-agent-how-lua-is-helping-african-businesses-move-ai-from-pilots-to/4cgx07z — The anatomy of an agent: Lua moves AI from pilots to production [2] https://www.passionateinmarketing.com/cygnet-one-launches-agentic-ai-platform-for-accounts-payable — Cygnet.One launches agentic AI platform for accounts payable