AI & automation
AI agents for business: choose the job before the platform
Most advice about AI agents starts with a list of tools. A more useful starting point is the job: a narrow, repeatable piece of work with clear inputs, a measurable result, and a person who owns the outcome. Choose that first, then decide how much the agent is allowed to do on its own.
The short answer
An AI agent is worth deploying when the task is frequent, bounded, and checkable, and when a mistake can be caught before it becomes expensive. Start with a narrow agent that suggests or drafts, measure its quality against how the work is done today, and raise its autonomy only when the evidence supports it. If a fixed sequence of steps can do the job, you may not need an agent at all.
What makes something an agent
The word is used loosely, often for products that are simply chatbots or rule-based automations with a new label. Anthropic’s engineering guide to building agents draws a practical line. In a workflow, models and tools are orchestrated through predefined code paths. An agent is a system where the model decides its own next steps and which tools to use. The difference matters because it changes who controls the sequence of actions—and therefore where things can go wrong.
| Approach | Who decides the next step | Typical job | Main risk |
|---|---|---|---|
| Rule-based automation | Rules written by people | Moving data between systems when a trigger fires | Breaks quietly when inputs change |
| Assistant or chatbot | The person using it | Answering questions, drafting text | A wrong answer accepted without checking |
| LLM workflow | Predefined steps; the model handles language-heavy steps | Classifying tickets, extracting fields from documents | Errors inside a single step |
| AI agent | The model, within the permissions it is given | Resolving a request end to end, researching and updating records | Errors that compound across steps and actions taken with real permissions |
We recommend finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.
Anthropic, “Building effective agents” (2024)
Why many agent projects stall
Adoption is real, but uneven. In McKinsey’s State of AI 2026 survey (1,719 respondents, fielded May–June 2026), about two in ten respondents said their organizations had reached the scaling phase with AI agents. Among large organizations the share was 40%, up from 27% a year earlier, while among smaller organizations it stayed flat at 22%. Scaling was most often reported in IT, knowledge management, and software engineering.
Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. The same release warns about “agent washing”—existing products rebranded as agents—and estimates that only about 130 of the thousands of vendors in the category offer real agentic capabilities.
Among teams already building agents, quality is the main obstacle. In LangChain’s State of Agent Engineering survey of 1,340 practitioners, 32% named quality as the biggest barrier to production, ahead of latency at 20%. The sample leans toward teams that are already invested in agents, but the pattern is useful: the hard part is reliable output, not access to a model.
How to choose the first job
A good first job for an agent is unglamorous. Use these criteria before looking at any platform:
- It happens often enough that a few minutes saved per case adds up.
- The inputs are accessible, and you are allowed to use them—including customer and personal data.
- A good result can be described and checked: a correct field, a resolved request, an accurate summary.
- Mistakes are recoverable: a draft can be corrected, a ticket reopened, a record restored.
- A named person owns the outcome and the exceptions.
- You know the baseline: how long the task takes today, how often it goes wrong, and what it costs.
Jobs that tend to disappoint have vague goals (“handle our marketing”), irreversible actions early on (payments, deletions, contractual commitments), scattered or undocumented inputs, or nobody responsible for exceptions. The same selection logic applies to automation in general, not only to agents.
Treat autonomy as a dial, not a switch
“Agent or no agent” is the wrong question. The better one is how much the system may do without a person. The OWASP Top 10 for Agentic Applications frames this as a principle of least agency: give an agent only the autonomy that a bounded task requires. In practice, that becomes a ladder.
| Level | What the agent does | Human role | Good fit | Evidence to move up |
|---|---|---|---|---|
| 1. Suggest | Recommends an answer or next step | Decides and acts | New or sensitive processes | Suggestions are usually accepted without major edits |
| 2. Draft | Prepares the output: a reply, a record update, a report | Reviews and sends | Customer replies, research summaries | Low correction rate and known failure patterns |
| 3. Act with approval | Executes specific actions after sign-off | Approves or rejects each action | CRM updates, refunds below a limit, scheduling | Approvals are rarely overridden; an audit trail exists |
| 4. Act within limits | Acts alone inside defined permissions and thresholds; escalates exceptions | Monitors, reviews samples, handles escalations | High-volume, low-risk requests | Stable quality over time; monitoring and rollback have been tested |
Each level should be earned with evidence from the previous one. Anthropic’s guide recommends extensive testing in sandboxed environments with appropriate guardrails, and notes that agents can pause for human feedback at checkpoints. Human review is also common in practice: 59.8% of LangChain’s respondents use it for nuanced or high-stakes situations.
Examples by team
| Team | Narrow first job | Starting level | What to measure |
|---|---|---|---|
| Customer support | Answer order-status and policy questions from your help centre and order data; escalate everything else | Draft, then act within limits for simple cases | Resolution rate, reopen rate, satisfaction on AI-handled cases, escalation quality |
| Sales | Research inbound leads and pre-fill CRM fields | Draft | Time per lead, field accuracy, how often reps keep the result |
| Operations | Extract data from invoices or forms and prepare records for entry | Act with approval | Extraction accuracy, exception rate, processing time |
| Internal knowledge | Answer staff questions from internal documents with links to the source | Suggest | Accuracy on a test set, share of answers with a valid source |
| Marketing | Turn an approved announcement into drafts for each channel | Draft | Editing effort, time saved, issues caught in review |
These are starting points, not recommendations for a particular product. Customer service is the most common agent use case in the LangChain survey (26.5% of respondents), followed by research and data analysis (24.4%).
A public lesson from customer service
Klarna is one of the most cited examples. In February 2024 the company said its AI assistant handled about two-thirds of customer service chats in its first month—2.3 million conversations. By May 2025 it had started recruiting human customer service staff again, and its CEO told Bloomberg that a focus on cost had come at the expense of quality, as reported by CX Dive. The assistant still handled most inquiries; what changed was the balance between automation and people.
The lesson is not that agents fail in support. It is that the level of autonomy was driven by a cost target rather than earned through quality evidence, and that customers still needed a reliable path to a person.
What agents cost, and how to judge the return
Pricing models differ: per user, per conversation, per resolved outcome, or consumption credits. As examples of published list prices, Intercom’s Fin charges $0.99 per outcome, where a resolution means no further help is requested after Fin’s last answer. Salesforce Agentforce is listed at $2 per conversation for customer-facing agents, or $500 per 100,000 Flex Credits, with per-user add-ons from $125 per user per month. Prices were last checked on 19 October 2026 and change often.
The licence is rarely the whole cost. Integration, data preparation, evaluation, monitoring, review time, and exception handling usually decide whether the numbers work. Anthropic’s guide also notes that agentic systems often trade latency and cost for better task performance—so the extra capability has to be worth paying for.
What to measure in a pilot
- Baseline time and cost per case before the pilot starts.
- Quality against a reviewed sample: accuracy, correction rate, reopen or complaint rate.
- Share of cases handled end to end versus escalated.
- Human time still required for review and exceptions.
- Total running cost per handled case, including platform, model usage, and review time.
- A decision rule agreed in advance: extend, redesign, or stop.
Common mistakes
- Starting from a platform demo instead of a job and a baseline.
- Granting broad permissions early. OWASP lists identity and privilege abuse and tool misuse among the top risks for agentic applications.
- Allowing irreversible actions before the agent has proven itself on reversible ones.
- Judging quality by impressions instead of a test set of real cases.
- Removing the human path for customers who need it.
- Paying for “agentic” features that are rebranded automation. Ask what the system actually decides on its own.
Checklist before you start
- Name one job and the person who owns its outcome.
- Measure how it is done today: time, cost, and error rate.
- Confirm data access and permissions.
- Choose the starting autonomy level and the evidence needed to move up.
- Build a test set of real cases, including the difficult ones.
- Run a time-boxed pilot with human review.
- Decide based on the results: extend, redesign, or stop.
If you want help defining the first job, the autonomy level, and a pilot you can measure, see our AI and automation work. When the use case itself is still unclear, a short product discovery phase is usually the faster route.
FAQ
Questions teams ask.
What are AI agents best at?
Frequent, bounded tasks with accessible inputs and checkable outputs, especially when the steps vary from case to case—support requests, research and data gathering, and document processing. They are weaker where goals are vague or mistakes cannot be undone.
Is ChatGPT an agent?
ChatGPT is primarily an assistant: it answers and the person decides what to do. Some AI products, including features in ChatGPT, can now take actions with tools, which is agent behaviour. For a business, the useful question is what the system decides and does on its own, and with which permissions.
How much do AI agents cost?
It depends on the pricing model. Published list prices include about $1 per resolved support conversation or $2 per conversation for some customer service agents. Integration, data preparation, monitoring, and human review are usually additional and often larger than the licence.
Do we need a custom-built agent?
Not always. Many jobs can be handled by agent features in tools you already use or by a simpler workflow. A custom agent makes sense when the job depends on your own data, systems, or rules that off-the-shelf products cannot reach.