# Build or buy? A guide for AI working groups in corporate finance

> Should a corporate finance firm build its own AI layer on Claude or buy one? What building really takes, who takes it from 90 to 99%, a 3-year cost of ownership, a factual comparison and how to use Mindro with Claude.

URL: https://www.mindro.co/build-vs-buy
Last updated: 10 October 2026

**Short answer:** "Build or buy" is the choice between a firm building its own AI layer, usually on top of Claude, and buying one. The AI layer is the plumbing under the agents: security, EU infrastructure, a semantic layer over the firm's own data, evaluations, connectors and an audit trail. Mindro is that layer for corporate finance firms, and it also brings a library of proven, ready-to-use agents and agentic workflows (the IM desk, the equity story desk, buyer matching, document classification and pitch preparation) that work off the shelf, so your firm starts with a head start instead of from zero. Building it yourself forfeits that kickstart. Your AI working group keeps Claude and builds the part that makes your firm different, the agents and workflows, on a foundation that is governed from day one. In our 3-year estimate for a 25-user firm, building the same layer yourself costs about three times as much, and the first live mandate comes 6 to 9 months later.

## What the AI working group gets right

Almost every corporate finance firm now has an AI working group: a partner sponsor, a team that meets weekly or monthly, and a mandate to move the firm forward. The group is right on the things that matter:

- **Experiment and move fast.** Trying tools on real work is how a firm learns what AI can do for it.
- **Keep the tools people like.** If your team is happy with Claude, keep it. Taking away tools people want slows adoption.
- **Build on your own expertise.** What makes your firm different is its deal history and its way of working, so the AI should be built around those.
- **Don't fall behind.** Three years of trying generic tools has not changed how most firms work. The groups that turn experiments into daily practice now pull ahead.

## From adoption to impact: three horizons

McKinsey's article ["From adoption to impact: Three horizons of AI transformation"](https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/from-adoption-to-impact-three-horizons-of-ai-transformation) (July 2026, a global survey of 750 employees and leaders, February to April 2026) describes three horizons: **enablement**, where employees get general-purpose AI tools for parts of their jobs; **automation**, where AI runs and improves cross-functional workflows at scale; and **reinvention**, where roles, workflows and operating models are redesigned. According to McKinsey:

- Only 11% of the leaders surveyed say their organisation is in the reinvention horizon.
- Leaders reporting meaningful enterprise value: 13% in enablement, 24% in automation and 48% in reinvention.
- 70% of respondents feel personally ready to use AI; only 27% of leaders think their organisation is ready to make the shifts needed.
- In the enablement horizon, leaders are 5.3 times more likely to report enterprise value capture when workflows are redesigned.

From our conversations with more than 100 M&A leaders in Europe: most corporate finance firms are still in enablement, and automation and reinvention are not yet on the agenda for most. Partner interest is there, but a strategy and a proven roadmap are often missing. Typically around 5% of a firm experiments happily with tools, skills, code and agents, and struggles to scale that firm-wide. Adoption is not transformation; moving on is a leadership question rather than a matter of personal readiness. All of Mindro's successful engagements involve partner-level commitment, and partners need new management skills: running a system of people and agents.

Mindro takes a firm from enablement to automation, with governed, evidenced deal workflows for IM creation, pitch preparation and buyer matching, and helps it prepare for reinvention, with AI strategy, a 12-month roadmap and change management included. The working group's experiments are not stopped but channelled into firm-wide capability through the agent hackathon, power-user sessions and desks you can use from Claude. [Our services](https://www.mindro.co/services)

## What building it yourself really takes

A demo on Claude is a weekend. A layer the firm can run live mandates on is a different project:

| Building block | What it involves |
|---|---|
| Security and compliance | Pen test, DPIA, EU residency set-up, single sign-on, data classification per tool |
| EU infrastructure | Vector search, a database, hosting and monitoring in Europe |
| Semantic layer | Structuring the firm's deal history and documents so AI can use them, and redoing it at every schema change |
| Evaluations | Testing that answers and figures stay right as models and prompts change |
| Connectors | SharePoint, CRM, email, meeting notetakers and data providers |
| Audit trail | Logging every data access by user, tool and time |

The people who volunteer for this are usually the firm's sharpest juniors. They are finance professionals, not engineers, and they do it alongside live mandates. The risk is not that they can't build an agent; it is that the firm's key people spend months iterating on plumbing instead of on deals.

## The last 10% is the hard part: who is your engineer?

A widely known engineering principle: getting an AI workflow from 0 to 90% reliability takes about as much effort as getting it from 90 to 99%. Proofs of concept typically reach roughly 80 to 90%, which is why the first demo on Claude feels so promising. M&A work needs near-99%, because everything must be evidenced.

So the real question for an AI working group is not "can we build it?" but "who is our engineer?": the person with the time and the skills to take the workflow from 90 to 99%, alongside live mandates. If you can name them, building may work for you. If you can't, a do-it-yourself build tends to stall at "reasonable". With Mindro, that engineering is included: forward-deployed engineers and data scientists work with your team, with engineering hours in every plan. [Our services](https://www.mindro.co/services#engineering)

## Reliability: "reasonable" is not enough for M&A

Evidenced M&A work needs output that is traceable and right, not just plausible. A self-made Claude skill can give reasonable results; an information memorandum needs every figure evidenced and checked. Mindro is built for that standard: at every stage, up to the final IM, one click takes you from any figure back to its source document. This end-to-end, one-click traceability is in use today on live mandates at corporate finance firms. [How Mindro creates an IM](https://www.mindro.co/information-memorandum)

## 3-year cost of ownership

| Year 1 | Do it yourself | Mindro |
|---|---|---|
| Build / platform | 1 FTE senior AI/data engineer, fully loaded: €120,000 (E) | Subscription €36,000 + setup €3,000 (F) |
| Security & compliance | Pen test, DPIA, EU residency set-up, SSO: €12,000 (E) | Included |
| Cloud infrastructure (EU) | Vector search, Postgres, hosting, monitoring: €15,000 (E) | Included |
| Semantic layer | Inside the engineer's time; re-done at every schema change | Included |
| Model tokens | €18,000 (E): one provider, no routing, no caps | €10,000 (E): routed across models, caps per user and project |
| Development hours for agents | Buy 48 h at €150: €7,200 (E) | 4 h per month included: 48 h per year |
| Tooling & licences | Observability, evals, connector SDKs: €5,000 (E) | Included |
| Your team's time | ~200 h specifying, testing, change management: €24,000 (E) | ~30 h: Mindro configures, trains and runs office hours: €3,600 (E) |
| **Year 1 total** | **≈ €201,000** | **≈ €52,600** |
| Years 2 and 3, per year | ≈ €114,000: 0.5 FTE maintenance, cloud, tokens, security refresh | ≈ €46,000: subscription and tokens |
| **3-year total cost of ownership** | **≈ €429,000** | **≈ €144,600** |
| First live mandate | Month 6 to 9 for an MVP; 9 to 18 for SSO, audit trail, connectors and agents (E) | Week 3 to 4 |

> Assumptions: 25-user boutique (10 analysts/associates), Advisory plan, blended internal cost €120/h. F = fact (Mindro pricing), E = estimate. Senior ML engineer NL ~€114k base (Jobmentis 2026). MIT NANDA 2025: vendor-built tools succeed ~2x as often as internal builds. Advisory plan: €3,000/month for up to 50 users, including data and AI engineering 4 hours/month; see [pricing](https://www.mindro.co/pricing).

## How the options compare

| Criteria | Enterprise LLMs (Claude, ChatGPT, Microsoft Copilot) | Rogo / Hebbia | Mindro | Build it yourself (internal team or consultancy) |
|---|---|---|---|---|
| Data residency and NDA compliance | Enterprise plans do not train on business data. EU options vary by vendor and plan: ChatGPT Enterprise offers storage in Europe; Claude offers EU regions through AWS Bedrock and Google Cloud Vertex AI; Microsoft 365 Copilot falls under the EU Data Boundary for EU customers, with Anthropic models excluded | Rogo: siloed customer environments; EU data residency not stated publicly. Hebbia: US and EU regional processing for storage and inference; dedicated tenant available | SaaS run fully in Europe; EU data residency guaranteed. TLS 1.2+ in transit, AES-256 at rest. No training on your data | Full control, provided the firm builds and runs security, hosting and monitoring itself |
| Firm-level governance | Set per tool and plan; Microsoft 365 Copilot honours Microsoft 365 permissions and sensitivity labels | Rogo: records scoped to who may see them; MNPI kept in a walled-off section. Hebbia: siloed environments; customers can see who accessed their data and when | Company realm: all confidential data in one governed environment, every data access logged by user, tool and time; access per connected tool configured per client | Configurable; the firm's own ongoing responsibility |
| Institutional knowledge | Retrieval over files users upload or connect, for example Microsoft 365 content a user can access | Rogo: custom agents can encode a firm's templates. Hebbia: draws on the firm's deal history, internal systems and external feeds | Firm brain built from the firm's own deal history and documents; owned by the firm, portable by contract | Possible with a vector database and a semantic layer the team builds and maintains |
| AI models | Each vendor's own models; Microsoft 365 Copilot uses OpenAI and Anthropic models plus models hosted by Microsoft | Rogo: models from OpenAI, Anthropic and Google. Hebbia: not stated on the pages cited | Orchestrates Claude, Gemini and Mistral; an AWS-based setup on request also allows OpenAI models. Claude and other tools can connect in a configured setup | Any model the team integrates and maintains |
| Implementation and support | Vendor support per plan; configuration and change management by the firm or a partner | Rogo: not stated on the pages cited. Hebbia: its team works with clients to index their deal history and processes | Included in every plan: training, workshop, agent hackathon, power-user sessions, office hours and data and AI engineering hours | Internal time or consultancy hours |
| Time to first live mandate | General tasks from day one; firm-specific workflows depend on what the firm builds | Not stated publicly | 2 to 3 weeks to deployment; first live mandate in week 3 to 4 | Estimated month 6 to 9 for an MVP (see the cost table) |
| Pricing | Per-user licences; API usage billed separately for custom builds | Not published | Advisory plan €3,000/month for up to 50 users; model tokens estimated at about €10,000 a year for a 25-user firm | Estimated ≈ €201,000 in year 1 and ≈ €114,000 a year after |

> Based on public information as of October 2026, taken from the vendors' own pages: openai.com/index/introducing-data-residency-in-europe, claude.com/regional-compliance, learn.microsoft.com (Microsoft 365 Copilot privacy and data protection), rogo.com, rogo.com/security, rogo.com/platform/intelligence, hebbia.com/security and hebbia.com/blog/introducing-matrix-2-0. Where something is not stated there, we say so. Mindro is one option among these; see also [all comparisons](https://www.mindro.co/compare).

## Use Mindro with Claude

You don't have to choose between Mindro and Claude.

- **Keep Claude.** In a connected setup, Claude can be connected to Mindro for the data your firm agrees it may use, while mandate data stays in Mindro. [How the company realm works](https://www.mindro.co/company-realm)
- **Use Mindro's desks from Claude.** In a connected setup, MCP and API connections are configured per client on request, so Claude and other approved tools can call Mindro's desks: the IM desk, the equity story desk, the buyer matching desk and the document classification desk. [Setups for institutions](https://www.mindro.co/institutional)
- **Build the fun part, not the plumbing.** Your working group builds agents and workflows on top of the firm brain. Security, EU infrastructure, the semantic layer, evaluations, connectors and the audit trail are already there.
- **Get help building.** An agent mini hackathon, power-user one-to-one sessions and data and AI engineering hours are included in every plan. [Services](https://www.mindro.co/services)

## For the partner

- **Happy people.** Your team keeps the tools it likes and gets to build, with support.
- **Real adoption.** Training, office hours and refreshers turn experiments into daily practice.
- **Not falling behind.** Your firm's own expertise becomes an AI capability that compounds with every mandate.
- **Lower risk.** Data stays in Europe, all confidential data is processed in one governed environment, and every data access is logged.
- **Speed.** The first live mandate in week 3 to 4.


### Summary for your AI working group

- Keep Claude. In a configured setup, it connects to Mindro for the data your firm agrees it may use, and Mindro's desks can be used from Claude.
- Start with a head start: ready-to-use agents and workflows (IM desk, equity story desk, buyer matching, document classification, pitch preparation) work off the shelf.
- Ask "who is our engineer?": getting from 90 to 99% reliability takes as much effort as getting to 90%, and with Mindro that engineering is included.
- Build the agents and workflows that make our firm different; don't build the plumbing (security, EU infrastructure, semantic layer, evaluations, connectors, audit trail).
- Evidenced M&A work needs traceable output: in Mindro, one click from any figure to its source document, up to the final IM.
- Estimated 3-year cost for a 25-user firm: ≈ €144,600 with Mindro vs ≈ €429,000 to build it ourselves.
- First live mandate in week 3 to 4 instead of month 6 to 9.
- Move from enablement to automation, in the terms of McKinsey's three horizons of AI transformation, with partner-level commitment.
- Read the full guide: https://www.mindro.co/build-vs-buy


## Frequently asked questions

**Our team built a proof of concept on Claude. Why not finish it ourselves?**
Because the last stretch is the hard part. Getting an AI workflow from 0 to 90% reliability takes about as much effort as getting it from 90 to 99%, a widely known engineering principle, and M&A work needs near-99% because everything must be evidenced. Ask who your engineer is: the person with the time and skills to take it from 90 to 99%. With Mindro, forward-deployed engineers and engineering hours in every plan do that work with your team, and your proof of concept can live on as an agent on top of the firm brain.

**Can we build this ourselves on Claude?**
You can build agents on Claude, and your working group should. Running live mandates also needs security, EU infrastructure, a semantic layer over your deal history, evaluations, connectors and an audit trail. In our estimate for a 25-user firm, building that layer costs about €429,000 over three years against about €144,600 with Mindro, and the first live mandate comes 6 to 9 months later.

**We already use Claude, why Mindro?**
Keep using it. Mindro adds the firm brain built from your own deal history, the company realm that keeps confidential data in one governed, auditable environment, and desks for IM creation, the equity story, buyer matching and document classification. In a configured setup, Claude can connect to Mindro for the data your firm agrees it may use, and Mindro's desks can be used from Claude.

**Will Mindro limit our team's experimentation?**
No. Your team keeps building agents and workflows, now on top of the firm brain, with an agent mini hackathon, power-user sessions and data and AI engineering hours included. The company realm only sets which data each tool may use.

**How long until our first live mandate?**
Deployment takes 2 to 3 weeks; the first live mandate is in week 3 to 4.

**Do we need our own developers?**
No. Data and AI engineering by Mindro's forward-deployed engineers and data scientists is included in every plan.

[Services](https://www.mindro.co/services) · [Setups for institutions](https://www.mindro.co/institutional) · [Mindro vs generic AI assistants](https://www.mindro.co/compare/mindro-vs-generic-ai) · [Book a briefing](https://www.mindro.co/contact)