The AI Mix.

Meet the pieces

The language of AI mixes together products, capabilities and technical plumbing. These are the distinctions that matter when deciding what to invest in.

These pieces can arrive inside software you already use. You do not need to purchase a separate platform for every row.

Assistants: give people a useful head start

An assistant helps a person work. The person supplies direction, examines the result and decides what happens next.

In our customer enquiry, an assistant could summarise the request, identify missing information and draft a reply. A salesperson checks the commercial details before sending it.

This is a useful starting point for work that involves writing, reading or analysis, particularly while a team is learning where AI performs well.

The important measure is the time and effort required to reach an acceptable result.

Workflows: make repeatable work repeatable

A workflow follows a route designed in advance. AI can handle flexible inputs within that route. For a straightforward quote request, the process might be:

1. AIExtract the requested product and quantity.
2. SoftwareCheck that the required details are present.
3. SoftwareRetrieve the approved price and stock position.
4. SoftwareCalculate the quote using business rules.
5. PersonSend it to the salesperson for approval.

The model interprets the email. Ordinary software performs the calculation. A person approves the commitment.

That division matters. Clear rules, exact calculations and database updates often belong in conventional software even when AI is part of the surrounding process.

Agents: handle the investigation

An agent becomes useful when the next step depends on what it discovers.

Suppose the requested product is unavailable. The system may need to check alternative products, compare specifications, inspect previous purchases and identify which unanswered question to take back to the customer.

An agent can choose among those steps within its instructions and permissions. Its design needs more than a goal: it needs suitable tools, reliable context, limits and a stopping condition.

For every agent, specify

• What result it owns.
• Which information it can read and which actions it can take.
• How success will be checked.
• When it must stop or ask for help.
• Who is responsible when the task fails.

Sub-agents: delegate only when delegation earns its keep

Imagine a complex tender involving several product categories and countries. A lead agent could delegate separate research tasks, then combine the findings into a recommendation.

That can help when the work is independently researchable or benefits from distinct expertise. It also creates additional work: assignments must be clear, findings must be reconciled and contradictions must be resolved.

Three agents agreeing does not establish that an answer is correct. They may share the same missing information or mistaken assumption.

Start by testing whether one agent with well-designed tools can complete the job. Add sub-agents when they produce a measurable improvement in quality, speed or manageable task scope.

Separate agents also need separately enforced access controls if their permissions are meant to differ. A role description alone is not an access boundary.

MCP: connect the intelligence to the business

MCP stands for Model Context Protocol. It provides a standard way for AI applications to access exposed tools, resources and reusable prompts.

In our example, an MCP server could make capabilities such as “look up customer”, “check stock” and “create draft quote” available to an AI application. Those capabilities may use existing business APIs underneath.

MCP is one connection option alongside native integrations and direct APIs. Its usefulness depends on the systems and applications you need to connect. The business questions remain concrete:

• Which actions are exposed?
• Whose identity is used?
• What information is returned?
• Where are permissions enforced?

A connection makes an action available. The surrounding system determines whether that action is appropriate and authorised.

Give AI the right knowledge, and a reliable record of progress

A model's general knowledge does not establish your current prices, customer terms or latest operating procedures.

Treat “memory” as a design decision. Decide what is retained, for how long, who can access it and how mistakes are corrected.

Keep forecasting and optimisation in the picture

Generative AI is only part of the business AI mix. A business may also need demand forecasting, fraud detection, recommendations, visual inspection or scheduling optimisation. Those needs may be better served by statistical models, specialist machine learning or optimisation software.

Different methods can work together. The method should follow the problem.

Choose authority deliberately

Reading a record, suggesting a decision and executing a transaction carry different consequences.

1. ReadRetrieve and summarise informationExplain an order's status
2. RecommendPropose a next stepSuggest an alternative product
3. PrepareCreate something for reviewSave a draft quote
4. Act within limitsExecute an approved class of actionUpdate a permitted internal field
5. EscalateRefer cases outside its authorityRequest approval for unusual discounts

Choose authority per action. Avoid giving a whole agent unrestricted access merely because some of its tasks are routine.

Test with realistic cases, including missing data, conflicting instructions and unavailable systems. Monitor the outcomes after launch, and maintain a practical way to pause the system or hand work to a person.

Build the business case around completed work

Begin with one process that has a clear owner, a measurable problem and accessible information. Record the starting position, then compare the AI-assisted process using the same measures:

• Volume
• Handling time
• Waiting time
• Error rate
• Current cost

Include human review, rework, integrations and maintenance in the cost. Distinguish time released from cash saved: capacity only becomes a financial saving when it changes spending, or creates value when it is put to productive use.

For the quote process, useful measures include time to an approved quote, accuracy of pricing, staff handling time and the proportion of requests requiring correction.

A suggested rollout sequence

1. Understand

   Map the process and measure its baseline

   Before proceeding: Is the problem worth solving?
2. Trial

   Test on representative examples in a controlled setting

   Before proceeding: Is quality acceptable, including exceptions?
3. Pilot

   Use with a limited team and defined approvals

   Before proceeding: Does it improve the complete process?
4. Expand

   Increase scope while monitoring outcomes

   Before proceeding: Can the business support it reliably?

This is a sequence of decisions, not a promise about implementation time.

Buy capabilities that meet ordinary needs within existing software where possible. Consider custom development when the process is distinctive enough to justify the integration and maintenance effort.

Your next move

Return to that Monday morning enquiry. Decide which parts require language understanding, which follow fixed rules, which require investigation and which carry commercial authority. That gives you a practical AI mix:

Add sub-agents if the investigation becomes complex enough to justify them.

Choose one valuable piece of work. Define what good looks like. Give AI the tools and authority it needs to complete that work, and measure the result.