AI and Automation
What businesses should automate before adopting AI agents
· 6 min read · LT Lab
Agents are most useful when the systems around them are already connected. Here is where to start, and why the unglamorous work pays off first.
AI agents promise to take on multi-step work: reading a request, looking up context, updating systems and drafting a response. The promise is real. But in most organizations, the first obstacle is not the model — it is the environment the agent has to operate in.
An agent can only act through the tools it is given. If customer data lives in three places, approvals happen in email and reports are assembled by hand, the agent inherits all of that fragmentation.
1. Connect the systems of record
Before an agent can update a CRM or raise an invoice, those systems need reliable APIs and a shared understanding of core entities — customers, products, projects and orders. Integrating your CRM, accounting and operational systems is often the single highest-value step toward useful automation.
2. Automate the deterministic steps
Many workflow steps follow clear rules: when a deal is won, create the customer in finance; when an invoice is approved, schedule payment. These should be handled by conventional automation, which is cheaper, faster and more predictable than an AI model.
Reserving AI for the parts that need reading, judgement or flexibility keeps costs down and makes the whole system easier to trust.
3. Make exceptions visible
Every automated process eventually meets a case it cannot handle. Before introducing agents, make sure there is a place for exceptions to go — a queue, an owner and a way to resolve them. Agents work best when they can escalate confidently.
4. Decide what 'good' looks like
Define measurable outcomes before building: turnaround time, error rate, volume handled without intervention. These become the evaluation criteria for any agent you deploy.
- Which tasks consume the most skilled time today?
- Which of those follow clear rules, and which need judgement?
- Which systems must the automation read from and write to?
- Who reviews the output, and when?
The practical sequence
Integrate, automate the predictable, instrument the exceptions — then introduce agents where judgement is genuinely required. Organizations that follow this sequence tend to find that agents deliver value quickly, because the groundwork makes every action safe and observable.