Insights / Guide / Generative AI, copilots and agentic automation
GuideGenerative AI, copilots and agentic automation: the complete guide
Generative AI drafts and answers; agents plan and act. How to choose use cases, deploy copilots safely, automate documents and set limits that keep agents in check.
Generative AI can draft, summarise and answer questions. Agents go further: they plan a sequence of steps and act across your systems, within limits you set. Used well, both remove hours of manual effort and speed up decisions. Used carelessly, they create new kinds of risk, from confident wrong answers to customer data leaking into public tools to automated actions nobody can explain afterwards.
This guide draws together the briefings in our AI series. It explains the differences between the main types of AI tool, how to choose where to start, how to deploy copilots and document automation, and how to set the permissions, limits and human checks that make agents safe to use. It is written for business leaders, not data scientists.
Know what you are buying
Three kinds of technology are often described with the same words, but they behave very differently and need different controls.
- Process automation follows fixed steps: when an order arrives, copy these fields into the finance system. It is predictable and cheap to run, but breaks when inputs change.
- Generative AI creates content: a draft reply, a summary of a contract, an answer drawn from your policies. It handles messy inputs well but can be wrong in ways that look convincing.
- Agents decide what to do next. Given a goal such as resolving a late delivery, an agent may look up the order, check the carrier, email the customer and update the record. That autonomy is both the value and the risk.
Many good solutions combine all three. Where rules are clear, plain workflow automation is cheaper and more predictable. Where inputs are messy, such as scanned documents or free-text emails, AI can read and classify them before automation takes over. Agents earn their place for multi-step tasks that need judgement within clear boundaries.
Start with value, not tools
Pick use cases by business value, feasibility and risk across the whole value chain. Separate quick wins that save hours now from bigger bets that change how you compete, and fund them differently.
- List the tasks that consume the most hours each week and follow a pattern: answering the same questions, typing details from documents, preparing the same reports.
- For each, estimate the time spent, the error rate and how quickly you could tell whether an AI answer was right.
- Check the data: does the information exist in digital form, and can it be trusted?
- Rate the risk: does the task touch customers, money, personal data or decisions about people?
- Choose one or two use cases that score high on value and feasibility and low to medium on risk.
A good first copilot project is narrow, frequent and easy to check: a sales team searching product specifications, a finance team answering supplier queries about payment status, a support team drafting replies from approved policies. Measure the time saved on a small group before rolling out further.
Deploy assistants and copilots securely
Assistants over your own documents and role-based copilots for finance, sales, HR, support and operations can return many hours a week to teams. Two conditions decide whether they help or mislead.
First, your data decides the result. An assistant grounded in outdated policies, duplicate price lists or draft contracts will give confident wrong answers. Before launch, make sure the documents it reads are current, approved and well organised, with a way to retire old versions. Second, access must respect who can see what. A copilot that can read every file on the network may show a junior employee the board papers or salary data. Connect it through your existing permissions so each user sees only what they already could.
Security and privacy first
Decide what data may enter which tool, keep client and personal data out of public tools, and test for prompt injection and data leakage before launch. A short written rule set prevents most early problems.
Adoption is half the project. Tools that are not trusted or understood go unused. Train people in a few simple prompting habits: state the goal and the audience, give the context and an example of the style needed, then review and refine the output rather than starting again. Appoint a champion in each team who collects good examples and shares them.
Automate documents and routine decisions
Some of the clearest returns come from reading documents: invoices, purchase orders, contracts, delivery notes and emails. AI can extract what matters, check it against your records and route it, with exceptions going to a person. Look for processes with high volume, stable rules and digital or scannable inputs. Avoid processes that change every month or depend on judgement nobody can write down.
- Is the process stable enough that the rules will not change every few weeks?
- Are the inputs digital, or can they be scanned reliably?
- Is the benefit measurable in hours, errors or cycle time?
- Is there a named owner who will maintain the automation?
- Is there a clear exception route, with someone who sees and resolves failures quickly?
- Will volumes, handling time, error rates and exception rates be reviewed monthly?
If any answer is no, fix that first. Every automated process meets cases it cannot handle, and an automation that fails silently loses trust quickly. When suppliers, forms or rules change, the automation must change with them, which is why a named owner matters.
Set limits for agents
Agents need permissions, limits and an audit trail. Treat an agent like a new employee with a tightly scoped role: what it can see, what it can change, what it can spend, when it must ask and a record of everything it did. Good first candidates are multi-step tasks with clear goals and boundaries, such as chasing late shipments, reconciling payments, preparing routine reports or resolving common support requests.
- Define read and write access separately, and grant the minimum needed.
- Set limits on volume and value, such as how many emails per hour or the maximum credit note it can raise.
- Require approval for anything irreversible or customer-facing at first.
- Log every action with the reason the agent gave for it.
- Test on realistic cases, including unusual ones, before release.
- Monitor success rates, overrides, costs and incidents, and keep the ability to switch the agent off quickly.
A sensible start is to run the agent in suggestion mode: it proposes actions, people carry them out and the two are compared. Grant limited autonomy step by step as the record shows the proposals are sound.
Design approval gates that work
Human review is the most common safeguard and the most often hollow. Place people at decisions that are costly, hard to reverse, legally sensitive or about people, and let automation handle low-risk, high-volume steps with sample review.
Give reviewers what they need: the source, the reasoning and the confidence, enough time, and the authority to reject without pressure to hit throughput targets. Watch for rubber-stamping. If approval rates are close to total and review times are very short, the gate may be decorative. Rotate reviewers, audit samples and include test cases where the right answer is to reject. Capture every correction with a reason and use it to improve prompts, rules and data.
Our reviewers were approving almost everything in seconds. Once we showed the source documents beside each proposal, they started catching real errors.
Head of finance operations, retail client
A quick design test helps for every AI step: what is the worst realistic error, who would notice, how fast, and how would it be undone? If nobody would notice, add a gate.
Done well, this work returns hours to your teams, puts answers drawn from your own data in front of people in seconds, applies the same checks every time and keeps AI use within clear owners, limits and records. Each use case built on a connected, secure foundation makes the next one cheaper. A practical first step is to list the five most time-consuming repeated tasks in one department and score them for value, feasibility and risk this month.
Source: Intelliversal analysis, drawing on the briefings in this series.
