Build vs Buy AI Agents for ERP and CRM

Short answer

Buy when your process is standard and the vendor already connects to your ERP and CRM. Build when the process is your own and no product reaches your systems or holds your rules. Most mid-size companies combine the two: a bought platform for models and controls, and custom integration for the logic that is theirs.

An AI agent is software that reads from your systems, decides on a next step and takes it, sometimes without a person in the loop. Vendors now sell agents inside most large ERP and CRM products, and the build-vs-buy pages that rank are mostly written by those vendors. This guide is neutral. It is for a CTO, CIO or operations director at a company of 100 to 1,000 staff that runs an ERP and a CRM.

What can an AI agent do inside an ERP or CRM today?

An agent can read records, draft a change and, with permission, write it back. Typical tasks are sorting incoming requests, matching documents to orders, updating CRM records after a call, and preparing a draft reply or quote for a person to approve.

Many companies already run agents. In the 2025 Work Trend Index, 46% of leaders said their organisation uses agents to fully automate workstreams or business processes. 82% expect to use digital labour to expand their workforce in the next 12 to 18 months.

The value is growing. BCG estimates that agents gave about 17% of AI value in 2025, and expects 29% by 2028.

When should you buy, and when should you build?

Buy when the process is the same in most companies. Build when the process is what makes your company different, or when no product can reach your systems.

Question	Buy	Build	Combine
Is the process standard (for example, CRM data entry)?	Yes	No	Partly
Does the vendor connect to your ERP and CRM out of the box?	Yes	No	One of them
Does the agent need rules that only your company has?	No	Yes	Yes
Does your ERP run on a local or custom system?	Rarely fits	Often	Often
Do you have staff or a partner to maintain code?	Not needed	Required	For the custom part
Must data stay in a set location or system?	Check the vendor	You control it	You control the custom part
Do you need it live in weeks, not months?	Yes	No	Partly

Count the answers per column. If one column wins clearly, start there. If the answers split, combine: buy the model access, the agent runtime and the controls, and build only the integration and business logic that are yours.

The same trade-off applies to software in general. Our guide to custom software vs off-the-shelf covers it in more depth.

What does integration cost, and who maintains it?

Integration is usually the largest cost in an agent project, and maintenance never stops. The model is a monthly fee. The work is in the connections: authentication, data mapping, error handling, retries and logs for each system the agent touches.

Ask these questions before you sign or start:

Who updates the connection when the ERP or CRM vendor changes its API?
Who pays for model fees as volume grows?
Who fixes a wrong action at night, and how fast?
Can you export the agent’s rules and logs if you change vendor?

Konzept builds integrations and does not sell an agent platform. Our AI automation work connects a model to the ERP and CRM you already run, with review steps and logs. When the agent needs custom code around a local system, that is software development work. Plan the budget and the owner for both.

Which controls must an agent have before it acts alone?

An agent needs limits, a human check, logs and an off switch before it writes to a live system without approval. Few companies have these yet. In its Applied AI Index 2026, BCG found that 42% of companies expect to give AI agents real autonomy by 2030, but only 5% have the full set of controls in place today.

Controls checklist

Scope. A written list of the actions the agent may take, and the ones it may not.
Least access. Its own account, with the smallest set of permissions.
Limits. A cap on value, volume and rate per action and per day.
Human approval. A person approves any action above a set risk or value.
Logs. Every input, decision and action is stored and can be traced.
Off switch. One step stops the agent, and the manual process takes over.
Testing. A test set of real cases runs before each prompt or model change.
Owner. One named person answers for what the agent does.
Review. A monthly check of errors, overrides and cost.

Give an agent autonomy one step at a time. Start with drafts, then approvals, then low-value actions on its own.

How many people does an agent need?

An agent still needs people: an owner, reviewers and a technical team. Microsoft calls this the “human-agent ratio”, a new metric for how many agents one person can direct. Its 2025 report defines the ratio but gives no number, so set your own.

Start with one reviewer for one agent. Raise the ratio only when the error and override rates stay low for several weeks.

Staff also use AI on their own. The 2024 Work Trend Index found that 78% of AI users bring their own AI tools to work, and 80% at small and medium-sized companies. Give them an approved route, or your agent governance covers only part of the AI use in the company. Our AI enablement program guide covers training and policy.

What does the EU AI Act ask of a company that deploys agents?

The Act sets duties for “deployers” of AI systems, and the AI literacy duty already applies to them. Check whether your use of an agent makes you a deployer. This article is not legal advice.

Reach. Article 2 covers third-country providers and deployers when the AI output is used in the EU. A company outside the EU should check where its agent’s output goes.
AI literacy. Article 4 says deployers “shall take measures to support the development of AI literacy” of their staff, with no “specific level” set. It applies from 2 February 2025.
High-risk dates. The European Commission confirms the AI Omnibus moved Annex III high-risk rules to 2 December 2027, and Annex I rules to 2 August 2028. To check scope, compare each agent’s task with the Annex III list, for example employment decisions.

How do you start with one process and grow from there?

Start with one process that is frequent, rule-based and cheap to fix when wrong. Prove it, then add the next.

Pick the process and measure it today: volume, time per case, error rate.
Fill in the decision table above for that process only.
Set up the controls list before the agent writes to a live system.
Run it in draft mode for four weeks, with a person approving every action.
Compare the results with the baseline, then widen its permissions one step.

Our guide to moving an AI pilot to production has a 90-day plan for this stage.

If you have one process in mind, ask for a quote. Send the process, the ERP and CRM you run, and the monthly volume. We will tell you whether to build, buy or combine.

Sources

Microsoft, The 2025 Annual Work Trend Index: The Frontier Firm Is Born — 23 April 2025
BCG, The Widening AI Value Gap — 30 September 2025
BCG, Applied AI Index 2026 press release — 30 September 2026
Microsoft and LinkedIn, 2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part — 8 May 2024
EU AI Act, Article 2 (unofficial text) — Regulation (EU) 2024/1689
AI Act Service Desk, Article 4: AI literacy — checked 28 October 2026
European Commission, AI Act regulatory framework — updated 19 October 2026

FAQ

Is it cheaper to build or buy an AI agent?

It depends on your systems and your process. Buying is usually faster when the vendor already connects to your ERP and CRM. Building gives you control when your process or systems are unusual, but you pay to maintain the code. Compare the full running cost over several years, including integration, model fees and support, not only the start price.

Can an AI agent write directly to our ERP?

It can, if the ERP has an API and you give the agent an account with write access. Do not start there. Let the agent draft changes that a person approves, with limits, logs and an off switch in place. Give it direct write access for low-value actions only after its error rate stays low.

What is the human-agent ratio?

It is a metric Microsoft introduced in its 2025 Work Trend Index for how many AI agents one person can direct and review. Microsoft gives no target number. Set your own: start with one reviewer per agent, track errors and overrides, and raise the ratio only when the results stay stable.

Does the EU AI Act apply to AI agents we use internally?

Article 4 AI literacy applies to deployers, so first check whether your internal use makes you one. Article 2 also reaches companies outside the EU when the AI output is used in the EU. Whether the high-risk rules apply depends on the task, so compare it with Annex III. This is not legal advice.