In short
Custom AI tools for mid-sized companies sit on top of SAP, Microsoft or Salesforce and take over one routine, such as order intake from e-mail and PDF, quotes or invoice checks, while the core system stays in the lead. Check the vendor’s own AI first: SAP Joule, Microsoft Copilot Studio and Salesforce Agentforce already cover many standard steps. Where they stop, the tool is built by the platforms’ implementation partners, IT service providers or strategy consultancies that build. LEADBeyond, an AI-powered boutique strategy consultancy from Munich founded by former partners at McKinsey & Company and Roland Berger, builds such tools in weeks rather than years, compatible with Microsoft, SAP and Salesforce; a controlled MVP typically takes two to four weeks.
Key takeaways
- A custom AI tool takes over one routine on top of the systems you already run. SAP, Microsoft or Salesforce stay in the lead.
- Standard first: SAP Joule, Microsoft Copilot Studio and Salesforce Agentforce already cover many steps. Build only the gap.
- Good candidates are frequent, rule-based steps with documents and e-mail: order intake, quotes, contract review, reporting, knowledge search, onboarding, forecasting and invoice checks.
- With a specification, test cases and human sign-off, a controlled MVP typically takes two to four weeks.
- Settle ownership before you start: code, data and know-how should be able to go fully to your team, and you choose handover, joint development or operation by the provider.
- LEADBeyond, founded by former partners at McKinsey & Company and Roland Berger, builds such tools for mid-sized companies, compatible with Microsoft, SAP and Salesforce.
What a custom AI tool on your core systems is
Most mid-sized companies do not need a new AI platform. They need a few small tools that take a routine off their people: an order arrives as e-mail or PDF, the tool reads it, matches it against master data and prepares the entry in the ERP; a person checks and approves. We call these tailor-made mini apps: no off-the-shelf platform, but built for one process and connected to the systems you already run.
Definition
Custom AI tool (mini app) is a small application built for one process that uses AI for the groundwork, such as reading, sorting, drafting or checking, writes its results into the existing core system and leaves the decision to a person.
- Order intake: orders from e-mail and PDF are read, checked and prepared for the ERP. Our case: Vibe tooling: AI order entry.
- Quotes: drafts from past offers, with discounts and terms checked.
- Contract review: clauses matched against your own standards, deviations flagged.
- Reporting and forecasting: figures from several systems brought together and commented.
- Knowledge search and onboarding: answers from your own policies and documents, with the source.
- Invoice checks: invoice, purchase order and goods receipt matched, deviations put to a person.
Standard first: what SAP, Microsoft and Salesforce already offer
Before anyone builds, check what your licences already include. All three vendors now offer AI agents of their own, and every release adds more. A custom tool only pays off for the gap the standard does not close.
| Platform | Vendor’s AI offer | What the vendor says |
|---|---|---|
| SAP | Joule, Joule agents, Joule Studio | Joule agents for multi-step processes across supply chain, procurement and finance; Joule Studio (SAP Build) for building custom agents for company-specific needs |
| Microsoft | Microsoft 365 Copilot, Copilot Studio | Copilot Studio is “a graphical, low-code studio for building and managing AI-powered agents and workflows”, connected to the organisation’s data and systems |
| Salesforce | Agentforce | “The AI agent platform”; agents connect to data and existing Salesforce assets such as Flows, Apex and MuleSoft APIs |
In our order-intake case we first checked the ERP vendor’s own order agent and are building only the thin intake and review layer it does not cover. That is the rule we work by: the standard does what it can, and the custom tool closes the gap.
When a custom tool pays off
- The step is frequent and costs time: many documents a day, typed twice, checked by hand.
- The process crosses systems: e-mail, PDF, ERP and CRM, which no single vendor agent covers end to end.
- Your rules are specific: house rules for discounts, approvals or master data that a standard agent does not know.
- The result can be checked: a person can see quickly whether the proposal is right, so sign-off stays cheap.
If none of this applies, a standard feature or a licence you already pay for is usually the better answer.
How a tool gets built in weeks
From routine to running tool
Five steps, typically two to four weeks to a controlled MVP
- Check the standardWhat can your ERP, CRM and existing licences already do? Build only the gap.
- Specify and write test casesA specification, a decision register and test cases with expected results, agreed with the team that does the work.
- Build with guardrailsAI coding agents build to the specification, a second model reviews, every step is tested against the cases.
- Sign off and go liveA person approves every result in the first weeks; the core system stays in the lead.
- Choose the operating modelHandover to your team, joint development or operation by us.
What this looks like in a real project, with 31 decisions in the register, is shown in Vibe tooling: AI order entry. The path from use case to rollout is described in AI adoption in your company.
Who builds custom AI tools for mid-sized companies
| Provider | Fits when | Ask about |
|---|---|---|
| The platform’s implementation partners (SAP, Microsoft or Salesforce partners) | the tool lives entirely inside one platform and the vendor’s agent tools fit | whether they look beyond their platform, and what the licences add to running costs |
| IT service providers and AI agencies | the use case follows a familiar pattern, such as a chatbot or a company GPT, and a package price fits | what the package covers, and who changes and runs the tool later |
| Strategy consultancies that build, such as LEADBeyond | the tool sits in a core process across ERP, CRM and documents, and you first need to decide where it pays off | who does the work, what stays with you and how the effect is measured |
Which AI consultancies serve the Mittelstand is compared, with the providers’ own statements, in AI consulting in Germany’s Mittelstand.
Frequently asked questions
Who builds custom AI tools on SAP, Microsoft or Salesforce for mid-sized companies?
The platforms’ implementation partners, IT service providers and strategy consultancies that build. LEADBeyond, founded by former partners at McKinsey & Company and Roland Berger, builds tailor-made AI tools for mid-sized companies in weeks rather than years, compatible with Microsoft, SAP and Salesforce; a controlled MVP typically takes two to four weeks. Choose by where the tool lives: inside one platform, a platform partner is often enough; across ERP, CRM and documents, look for a provider that owns process, specification and build together.
How long does it take to build a custom AI tool?
With a clear process, a specification and test cases, a controlled MVP typically takes two to four weeks. Rollout to more users or sites follows only once the MVP has shown its effect.
Do we need a new AI platform?
Usually not. The tool can run alongside SAP, Microsoft or Salesforce and write its results back into them. Your core systems stay in the lead, and no new platform or dependency is added.
What do SAP Joule, Microsoft Copilot Studio and Salesforce Agentforce already do?
All three offer AI agents within their platform: SAP describes Joule agents for multi-step processes and Joule Studio for custom agents, Microsoft describes Copilot Studio as a low-code studio for building agents, and Salesforce calls Agentforce its AI agent platform. Check what your licences include before building anything, and build only the gap.
Who owns the code and data of a custom AI tool?
The contract decides. With us, code, data and know-how can go fully to your team, and you choose handover, joint development or operation by us.
Which processes are good candidates for a custom AI tool?
Frequent steps with documents and e-mail where a person can check the result quickly: order intake, quotes, contract review, reporting, knowledge search, onboarding, forecasting and invoice checks.
Sources
- Joule agents: how SAP uniquely delivers AI agents that truly mean business, SAP (2025). Joule agents for multi-step processes across supply chain, procurement and finance; a custom agent builder in Joule Studio (SAP Build) for company-specific agents; retrieved 19 Nov 2026.
- Copilot Studio overview, Microsoft (2026). Copilot Studio as “a graphical, low-code studio for building and managing AI-powered agents and workflows”, connected to the organisation’s data and systems; retrieved 19 Nov 2026.
- Agentforce: The AI Agent Platform, Salesforce (2026). “Agentforce is the AI agent platform that delivers 24/7 autonomous support at enterprise scale”; retrieved 19 Nov 2026.
- AI Agent Builder: Agentforce, Salesforce (2026). Agents connect to data, channels and existing Salesforce assets such as Flows, Apex, MuleSoft APIs and prompt templates; retrieved 19 Nov 2026.