What vendors offer, what buying vs. building means, and how to decide
AI agents now carry out real work inside businesses. A firm can buy an agent ready-made or build its own. This guide explains what AI agents are, what the providers offer, and how to make the buy-or-build decision, without wasting months or thousands.
AI agents are already changing how work is done. Unlike a chatbot, which answers a question and leaves the work to you, an agent carries out the work itself: it operates your systems and produces the result. How much the agent does on its own, and how closely a person stays involved, is yours to decide.
Firms are using agents to do real work and gain a real edge. Among the processes AI agents are delivering:
You no longer have to build an agent from scratch to get these results. Ready-made agents are now available in three ways: free templates published by the AI labs that you adapt, paid products built into enterprise software you may already run, and managed services that a provider hosts and runs for you. Buying an agent can save months of build time, though it brings trade-offs in control, fit and cost that are easy to miss.
This guide explains what these agents are, what each provider offers, when buying is the right move and when building is, and the work that remains yours either way.
A chatbot answers questions. You ask, it replies, and acting on the reply is still your job. Most people's first experience of AI works this way, which is exactly why agents are easy to underestimate.
AI agents are different in kind. AI agents are systems that can carry out business operations on their own: using your IT systems, weighing options, producing deliverables, coordinating with other agents, and performing judgement-based work that until now needed a person. You give an agent an objective in plain language, the way you would brief a colleague, and connect the agent to the relevant systems and data. From there the agent plans the steps and carries them out. Several agents can work together as an agentic system, capable of long-running, complex work that runs over hours rather than seconds.
The difference is clearest in a real process. Take onboarding a new client.
“What is our client onboarding procedure?”
It returns the steps, the checks and the policy, clearly explained. Carrying them out is still entirely your team's job.
“Onboard this client.”
The agent runs the know-your-customer checks, opens the accounts in your CRM and finance systems, drafts and sends the welcome communications, and escalates only the cases that need a human decision.
The chatbot describes the work. The agent does it.
How much an agent does on its own is a choice, and it runs along a scale. At the lower end, the agent supports a person who does the work. At the higher end, the agent runs the whole process and a person handles only the exceptions. The lower end is often a straightforward workflow with a person in the loop, and it is where a great deal of the early value lies. More autonomy is not automatically better: it raises both the value the agent can deliver and the oversight the agent needs.
The agent gathers and presents what a person needs, and the person makes the decision and does the work. This is the simplest and safest use, and often the most valuable place to start.
The agent proposes a specific action or answer, and a person decides whether to accept it. The judgement still rests with the person.
The agent carries out the task, but a person reviews and approves the result before it takes effect or reaches anyone else.
The agent runs the whole process on its own and refers only the exceptions to a person. This delivers the most, and it demands the most oversight, control and trust.
Choosing the right level for each task is one of the most important decisions you will make. The platforms give you the setting; deciding where each of your processes belongs, and building the oversight around it, is the part that is genuinely yours.
That shift, from describing work to doing it, shows up in three ways a leader can measure.
Time
Hours become minutes
Work that took a person hours is completed in minutes, freeing those hours for the judgement only people can provide.
Output
More output, same team
The agent handles the routine, repeatable work, so the team's capacity goes to the decisions worth a person's attention.
Cost & quality
Fewer errors, less rework
The same team clears more work with fewer mistakes, which is the case that ultimately gets signed off.
An agent is only as good as the job you point it at and the limits you set around it. Aimed at the wrong work, or allowed to act without boundaries, it makes mistakes quickly, confidently and at scale. Getting that judgement right matters more than which provider you choose, and it is what the rest of this guide is about.
A ready-made agent is one you obtain rather than build from scratch. The large AI providers and the major enterprise-software vendors now offer ready-made agents in three ways. The differences matter for cost, control and how much work is left to you.
Vendor wordmarks shown as placeholders; official logos to be added at build.
Option 1 · Free
The AI labs publish complete agent templates, free to take, that you tailor to your own conventions and run on a platform you pay for. Anthropic publishes a growing library for financial services, covering work such as pitchbooks, know-your-customer screening, the month-end close and valuation review,1 plus templates for healthcare and life sciences,2 and a full suite for legal teams, into which software such as Thomson Reuters CoCounsel can plug.3 OpenAI offers prebuilt templates the same way.4
Suits: firms with some technical capability, who finish, connect and operate the template themselves.
Option 2 · Paid
The enterprise vendors sell ready-made agents as paid features within the systems you may already run, configured rather than built. Salesforce offers pre-built Agentforce agents for sales, service, marketing and commerce.5 Microsoft provides them through its Agent Store and Copilot Studio.6 Google offers prebuilt agents through the Agent Garden in Gemini Enterprise.7
Suits: teams that want the fastest route, with enterprise controls included, and are content to work the way the product works.
Option 3 · Managed
The provider hosts and runs the agent on the provider's own platform, so the agent can work on a schedule or across a long task without a person watching each step. Anthropic's Managed Agents platform supplies the surrounding machinery a firm would otherwise build: limits on which systems the agent may touch, a secure store for credentials, and a complete audit trail of every action.8
Suits: work that must run unattended or at scale, with a full record for auditors. Less suitable where a person should approve every result, or the data should not leave your environment.
The question is no longer whether you can get an agent without building one, because you can. The real question is which of these three routes fits each job, and how much autonomy and oversight that job should have. The platforms give you the autonomy settings and the permission controls as features, but they do not decide where each of your processes should sit, or design the oversight around your business. Choosing the right job, connecting the agent, and setting and governing its level of autonomy remain yours. That work is the subject of the rest of this guide.
The useful question is not whether to buy or build in general. It is which jobs deserve a build at all. Engineering effort and senior attention are scarce. The firms that get value from agents spend that effort where an agent gives them an edge, and obtain the rest the quickest way they can. Three things should drive the choice for any given job.
The closer a job is to what every firm does, the more likely a product or template already does it well, and the less a custom build returns. The further it sits from the standard, shaped by how your firm specifically operates, the more a build earns its cost.
Some work cannot run on a third party's terms: regulated or sensitive data that must stay in your environment, or behaviour you cannot afford a vendor to change underneath you. Where control is not negotiable, that points to a template you run yourself, or a build, rather than a productised agent.
A productised agent ties your capability to the vendor's roadmap and pricing. A template ties you to a platform but leaves the configuration in your hands. A build ties up your own team. None is free of dependency; the real choice is which one you would rather live with.
A free template sits deliberately between buying and building. You take the vendor's design for nothing, which removes much of the construction, but you still own the work that decides the outcome: connecting it to your data, fitting it to your process, and governing what it does. A template lowers the cost of building. It does not remove the responsibility.
The table sets the three routes against the questions a leader actually asks, with the implication of each spelled out. We keep it current as the providers change, which is why it lives here and not in a document downloaded once.
Scroll the table sideways to see all three routes →
| The question you are asking | Buy a productised agentPaid, inside your software | Adapt a free templatePublished by an AI lab | Build bespokeIn-house or with a partner |
|---|---|---|---|
| What you get | A finished agent sold inside enterprise software you may already run, set up by configuration.You switch it on rather than build it. | A complete, working starting point published by an AI lab, which you adapt and connect to your data.You finish it, then run it. | A system designed around your exact work.Nothing is assumed; everything is yours to define. |
| Examples available now | Salesforce Agentforce; Microsoft Agent Store and Copilot Studio; Google Agent Garden. | Anthropic finance, healthcare and legal templates; OpenAI templates. | Built in-house, or with a delivery partner such as Serpin. |
| How you pay | A subscription or per-use fee.Predictable to budget, but the meter runs for as long as you use it, and the price is the vendor's to change. | The template is free; you pay for the platform that runs it.Low to start, though your adaptation and running costs are real and yours. | A larger up-front build cost, then the cost to run it.Highest to begin, but you are not paying a vendor a fee on top forever. |
| Time to a working result | Shortest; live in days.The main reason to buy when the job is standard. | Short to moderate, depending on how much you change.A strong head start, but the adaptation is real work. | Longest; designed and tested from scratch.The price of a system that fits exactly. |
| Fit to how you work | Whatever the product does.If your process differs, you change your process or accept the gap. | Shaped to your conventions, within the template's design.Good fit on the common parts, limits on the rest. | Exact.The system follows your process, not the other way round. |
| Control over data and behaviour | On the vendor's terms.The vendor can change how the agent works, raise the price or retire it, and you must check how your data is handled and whether it can be used to train their models. | Largely yours, on a platform you choose.You can keep regulated or sensitive data inside your own environment. | Entirely yours.No third party can alter behaviour, pricing or data handling without you. |
| Dependence on the provider | High.Your capability rises and falls with the vendor's roadmap, and switching later means redoing the integration. | Moderate.You rely on the underlying model and platform, but own the configuration and can change how it runs. | Low on any one vendor, high on your own team.The capability has to live in-house or with a partner you trust. |
| Competitive advantage | Limited.A competitor can buy the identical agent, so it raises your floor without setting you apart. | Some.The starting point is shared, but how you adapt and combine templates can be your own. | Greatest.The agent is yours alone, which is the case for building where the work is core to your edge. |
| Technical capability needed | Least.Configuration and administration within the vendor's tools. | Some.Enough skill to finish, connect and operate the template safely. | Most.Real engineering and design capability, in-house or through a partner. |
| What stays your responsibility | Choosing the right job, connecting the agent to your systems, governing what it may do, and running it.Buying removes the construction, not the accountability. | All of the above, plus finishing the template to fit.You own the result. | The whole system, end to end. |
Read the final row across the three columns and it barely changes. Choosing the right job, connecting the agent to your systems, governing what it is allowed to do, and paying to run and improve it remain yours whichever route you take. Buying changes how much of the construction you skip, not who is accountable for the result.
Acquiring or building the agentic system is only part of the task. Most of the effort, and most of the cost, lies in the elements around it. These apply equally whether you buy or build.
| The right use case | Select work where an agentic system delivers a clear saving in time or cost. The wrong use case undermines the entire investment. |
|---|---|
| Integration with your systems | The agentic system must connect to the software where your data and work already reside, such as your email, CRM, finance and document systems. |
| Fit with your processes | It must align with how work actually flows through the business, and with the points where it returns tasks to a person. |
| Your organisation and people | An agentic system changes how the organisation works, not only its technology. It can affect the ideal structure, processes, roles and skills around it, and your people need training and a reason to trust it, or they will work around it. |
| The level of autonomy | Decide, for each task, how much the agentic system does on its own and where a person stays in the loop. The platforms offer this as a setting, but choosing where each process should sit, and designing the oversight around it, is the harder and more valuable judgement. |
| Governance and oversight | Clear rules for what the agentic system may do, a record of its actions, and a named owner accountable when it gets something wrong. |
| Security and data protection | Control over what the agentic system can access and do, so that sensitive and regulated data remains secure and private. |
| Cost and return | The full cost to run and improve the agentic system over time, weighed against the time and money it saves. This is the figure that determines whether it is worthwhile. |
| Ongoing maintenance | Monitoring its output and correcting it as your business and the underlying AI models change. An agentic system left unattended becomes less reliable over time. |
Serpin's methods address each of these: the First AI Agent Playbook for selecting the use case, Agent Discovery and Design for systems, processes and people, and Bounded Agency for governance and oversight.
A clear sequence for moving from an initial idea to a decision you can defend, and on to an agentic system that works in practice.
Select a specific task where an agentic system can deliver a measurable saving in time or cost. The strongest candidates are high-volume, repeatable processes with a clear, checkable output, such as first-meeting research, document review or reconciliation.
A narrow, well-defined task is easier to scope, to measure and to govern than a broad ambition such as “use AI across the business”, and it produces a result you can build on.
Estimate the value in concrete terms: hours saved, costs reduced, errors avoided and revenue enabled. Set that against the full cost of delivery, including building or buying, integration, and running the agentic system over time.
Running cost warrants particular attention, because the economics differ from traditional software. Where conventional software is largely a fixed cost, an agentic system incurs a charge every time it runs, and that charge grows with both the volume of work and the complexity of each task. Design and architecture choices, such as which model handles each step and how much information is sent to it, have a substantial effect on this figure, and an agentic system that is successful but poorly engineered can become an expensive one. A realistic estimate of running cost, not only build cost, belongs in the business case.
Establish whether a ready-made product or a vendor template already performs the task to an acceptable standard. Where the work is common to many firms, buying or adapting is almost always faster and less expensive than building from scratch.
Use the comparison in section 4 to weigh the routes against your requirements, and reserve a custom build for the work that genuinely warrants the investment.
Confirm the agentic system can connect to the systems where your data and work reside, and that it fits the way the process actually runs, including the points at which it returns work to a person.
Test it against your security, data-protection and regulatory requirements, and judge whether the people who will rely on it are ready to adopt it. A capable agentic system that does not fit the organisation will not be used, and the cost of finding that out after launch is high.
Buy or adapt where a product fits the task well and the work is not central to your competitive position. This is the faster route to a result, and it frees your own effort for the work that sets you apart.
Build where the task is core to your advantage, where no product matches how you operate, or where your data and control requirements preclude a third party. Most organisations arrive at a considered mix of both over time, by deliberate choice rather than by accident.
Before the agentic system goes live, define what it is permitted to do, what is recorded, and who is accountable for its decisions. Decide where a person must review or approve its work, and set clear limits on the systems and data it can reach.
Treat this as a change project, not a technology project. An agentic system can reshape the ideal organisation structure, processes, roles and skills around it, so plan for those changes, not only the deployment. Brief and train the people affected, give them a clear route to raise problems, and lead the adoption deliberately. This is usually the difference between a pilot that impresses and a system that endures.
The work of building or configuring the agentic system now follows one of two paths, depending on the route you chose in stage 5.
Configure the product or template to your data and conventions, within the limits set in stage 6. Connect it to the systems it needs, map it onto your process, and tailor its outputs to your house standards.
Adapt rather than rebuild. Most of the value of buying comes from accepting the product's design where it is good enough and changing only what genuinely matters. Validate its output on real cases before anyone relies on it.
Design the agentic system around the task: the steps it takes, the systems and data it draws on, the decisions it may make alone, and the points where it must defer to a person. Engineer the controls and checks around the model in code, not in the prompts alone.
Build and test against real examples, measuring quality, cost and reliability before launch. Architect for running cost from the outset, since the design choices made here largely determine what the agentic system will cost to operate.
Begin with a limited, supervised rollout on real work, and compare the results against the forecast from stage 2: the time saved, the running cost, and the quality of the output. Expand only once the agentic system performs reliably.
An agentic system is not a one-off installation. Its accuracy and its cost drift as your business changes and as the underlying models are updated, so it needs a named owner and a regular review. Monitor its output for errors, track its running cost against budget, and refine the prompts, data and controls as you go. This ongoing operation is where many programmes quietly fail, and where a well-run one continues to compound its value.
We help firms bring AI agents into their work, whether you buy, build, or do some of both.
Book thirty minutes on a real job you have in mind. We will help you decide whether to buy or build it, and what it would take.
All sources verified against source on 18 June 2026.