Parts 1 and 2 covered what AI agents are and how they are built. This part covers what they are actually doing — in enterprise software, in developer tools, in customer service, in knowledge work, and in personal productivity. The goal is an honest assessment: where deployment is mature and well-evidenced, and where it remains promising but experimental.
The most important framing principle for this entire section: agent reliability in production is not the same as agent capability in a demo. Agents that perform impressively on a well-crafted demonstration task may fail unpredictably on real-world variants of the same task. We distinguish between these where the evidence is clear.
Software Development — The Most Mature Agent Domain
Software development was the first domain where AI agents achieved production-grade reliability for a simple reason: code has objective correctness criteria. It either runs, passes tests, and produces the expected output — or it does not. This makes agent evaluation measurable and agent progress trackable in a way that most other domains cannot match.
SWE-bench: the standard measure SWE-bench is the benchmark for software engineering agents: a dataset of real GitHub issues from production open-source repositories, evaluated by whether the agent's proposed fix passes the associated test suite. GPT-4's baseline performance: 1.74%. Devin (Cognition AI, 2024): 13.86%. The best 2025 systems: approximately 40–50%. This is genuine progress on genuinely hard tasks — and a clear signal that roughly half of well-specified software tasks still require human involvement. |
Agentic coding tools (Claude Code, Cursor Agent, Windsurf)
Terminal and IDE-integrated agents that can read entire codebases — not just the current file — understand cross-file architecture, implement features spanning multiple modules, write and run tests iteratively, refactor code with awareness of downstream effects, and fix build failures by reading error output and modifying code until it passes.
This is fundamentally different from autocomplete tools like the original GitHub Copilot. Autocomplete suggests the next line or function; agentic coding tools pursue a stated goal across an entire codebase over a multi-hour session, with full read and write access to all project files.
GitHub Copilot Workspace and PR automation
Takes a GitHub issue description as input, generates a natural-language implementation plan, scaffolds code changes across multiple files, runs the test suite, and proposes a pull request for human review. A multi-hour developer workflow — read the issue, understand the codebase, plan the changes, implement them, write tests, write the PR description — reduced to minutes of human review time.
Automated testing and QA
Agents that generate comprehensive test suites from specifications, run them against new code, identify failures, diagnose root causes in the call stack, and propose targeted fixes. Test coverage — historically limited by the time cost of writing tests — is one of the most tractable agent use cases because the success criterion (all tests pass) is unambiguous.
Legacy code modernisation
Translating COBOL and FORTRAN to modern languages, generating documentation for undocumented codebases, explaining what inherited code does without existing documentation. The agent reads the code, reasons about its behaviour, and produces explanations and modern equivalents — a task that would require a senior developer weeks of careful reading.
Knowledge Work and Research — Rapidly Maturing
Deep research agents
Given a research question, the agent searches the web and academic databases, retrieves and reads full papers and articles, synthesises findings across sources, identifies contradictions and gaps, and produces a structured report with citations. Products: Perplexity Deep Research, OpenAI Deep Research, Gemini Deep Research — all launched in 2024–2025.
The value proposition is clear: a research task that would take a human analyst several days of reading and synthesis takes the agent a few hours of compute. The limitation is equally clear: the agent cannot evaluate the credibility of sources as a domain expert would, may surface outdated or retracted findings, and produces outputs that require expert review before being acted upon.
Competitive intelligence monitoring
Agents that continuously monitor competitor websites, pricing pages, job postings, press releases, product announcement blogs, and earnings calls; synthesise changes into structured briefings; and alert when significant developments occur. This replaces the brittle combination of Google Alerts and manual analyst review with a system that reads full-page context and can recognise signal from noise.
Due diligence and document analysis
Agents that ingest a company's document data room — financial filings, contracts, board minutes, IP registrations — cross-reference claims against public records and news archives, flag inconsistencies, and produce structured due diligence reports. Law firms and private equity firms are among the earliest enterprise adopters, reporting significant time savings on first-pass document review.
Scientific literature synthesis
Agents monitoring ArXiv, PubMed, and journal feeds for new publications in a specified research domain, extracting key findings and methodologies, linking them to prior work in a maintained knowledge graph, and surfacing papers relevant to a researcher's specific interests within hours of publication. Replaces the researcher's need to manually scan feeds, while preserving the expert judgment required to evaluate and act on the findings.
Join our newsletter for regular updates on AI, digital marketing and growth!
Customer Service and Support — Mature in Bounded Deployments
Customer service illustrates both the maturity and the current limits of AI agents better than almost any other domain. The contrast with what preceded agents — keyword-matching chatbots that could answer a narrow set of scripted questions — is dramatic. The gap between current capability and what full autonomous resolution would require is equally instructive.
Beyond scripted chatbots: agents that actually resolve issues
A chatbot told "I need to return my order" looks up the return policy and explains the process. An agent with CRM, order management, and billing system access looks up the specific order, verifies it is within the return window, initiates the return label, schedules the refund upon receipt, and sends a confirmation — all without a human touching the ticket.
This is the distinction that matters: the chatbot answers questions about processes; the agent executes them. The agent's capability is bounded by the tools and permissions it has been given — it can only do what its connected systems allow.
Production deployments and reported performance
Sierra AI, Intercom Fin, and Salesforce Agentforce are purpose-built customer service agent platforms reporting resolution rates of 60–80% on Tier 1 and Tier 2 issues without human involvement. Human escalation paths are preserved for genuinely complex, ambiguous, or sensitive cases. The economic model: agents handle volume; human agents handle judgment.
Internal IT helpdesk is among the most mature deployments: password resets, software provisioning, VPN troubleshooting, device management, access request processing. These account for the majority of helpdesk ticket volume at most organisations, and they follow predictable patterns that agent systems handle reliably. Early adopters of AI-integrated service management (ServiceNow, Jira Service Management) report 40–60% ticket deflection rates.
Where human agents remain essential
Agents struggle with cases that require genuine judgment: customer situations that fall outside documented processes, emotionally sensitive interactions where tone and empathy matter as much as resolution, complex complaints involving multiple parties and disputed facts, and situations where the right resolution requires knowledge that is not in any connected system.
The well-designed customer service agent system does not try to replace human judgment on these cases — it routes them to humans with full context pre-populated, so the human agent can focus on the judgment and relationship work rather than information gathering.
Business Operations — Wide Variation in Maturity
Sales development (well-evidenced)
Agents that research prospect companies using recent news, job postings, funding announcements, and product changes; identify the relevant contact; write a personalised outreach message that references specific context from the research; schedule follow-up sequences; update CRM records after each interaction; and draft responses to inbound enquiries.
The value proposition is scale without proportionate headcount growth: each agent instance can manage hundreds of outreach sequences simultaneously, maintaining personalisation and contextual relevance that batch-email tools cannot achieve. The limitation is the same as all agent systems: errors compound, and an agent that misidentifies the relevant contact or misjudges the tone of an outreach email causes reputational harm that is harder to measure than a missed database field.
Document processing (well-evidenced for structured documents)
Ingesting large volumes of structured documents — invoices, contracts, forms, regulatory filings — extracting specified fields and clauses, validating against business rules, flagging exceptions, and writing structured outputs to downstream systems. For document types with consistent structure (standard invoice formats, employment agreement templates), agent accuracy is high and the labour savings are substantial.
For highly variable or jurisdiction-specific documents, accuracy degrades. The agent performs well on documents that resemble its training data; it struggles on novel formats and edge cases. Expert review of exceptions remains essential.
Financial reconciliation (early-stage deployments)
Agents that ingest transaction records from multiple systems — ERP, bank feeds, payment processors — identify mismatches, trace reconciliation chains, and produce structured exception reports for accountant review. The mechanical matching work that constitutes the majority of reconciliation volume is well-suited to agents. The judgment calls on ambiguous transactions, client communication, and tax strategy remain human work.
Procurement and vendor management (emerging)
Agents that gather quotes from multiple vendors based on specifications, compare against procurement policy, flag policy deviations, and summarise proposals into structured comparison formats for human decision-making. The human makes the vendor selection; the agent handles the information-gathering and comparison workflow. Early deployments are promising; production readiness varies significantly by industry and document complexity.
Personal Productivity — The Consumer Frontier
Email management
Agents that categorise incoming mail, draft replies in the user's voice for human approval, unsubscribe from mailing lists based on stated preferences, summarise long threads, and surface time-sensitive items. The pattern — agent drafts, human reviews and sends — is well-suited to the irreversibility concern that governs agent design: external email is an irreversible action and benefits from a human approval gate.
Calendar and scheduling coordination
Finding mutual availability across participants in multiple time zones, booking meeting rooms and video conference links, sending invitations with context briefings, preparing agenda documents, and managing rescheduling when conflicts arise. High friction, low cognitive value — a near-ideal agent use case. The actions are reversible (meetings can be cancelled), the task structure is well-defined, and the performance bar is clear: did the meeting get booked at a time that works for everyone?
Travel planning
Searching across flight and hotel options, filtering by stated preferences (airline loyalty programme, maximum layover time, budget ceiling, neighbourhood preference), comparing and completing bookings, building itineraries, and monitoring for disruptions or price changes. Consumer-facing implementations (Booking.com AI, Expedia AI, Google Travel) are early examples; the quality of current deployments varies significantly.
The maturity spectrum at a glance Mature and production-ready: Software development (coding, testing, QA), IT helpdesk automation, structured document processing, calendar coordination Established with important caveats: Customer service Tier 1-2, sales development outreach, competitive intelligence monitoring Promising, deployment-dependent: Deep research synthesis, due diligence, financial reconciliation, procurement workflows Early-stage, experimental: Personal productivity agents (email, travel), fully autonomous business process management |
Summary and What's Next
AI agents are already delivering measurable productivity improvements in software development, customer service, and document processing. The pattern is consistent across all mature deployments: agents handle volume, routine cases, and well-structured tasks; humans handle judgment, exceptions, and irreversible decisions.
The most common deployment mistake is overestimating agent reliability on novel or edge-case inputs and underestimating the cost of the errors that result. The second most common is underestimating the infrastructure required: tools, memory management, observability, and human oversight mechanisms all need to be designed, not assumed.
Part 4 — the final instalment in this series — covers why agents fail, in specific and actionable detail, and what safety and security measures are non-negotiable before production deployment.