QuantSpark
Get in touch

Choosing a forward-deployed AI partner

Three categories of forward-deployed AI partner. Choose deliberately

OpenAI and Anthropic have both walked into the consulting business. The forward-deployed engineer model that Palantir invented in 2003 is now the baseline.1 If you are buying it, you have three real options. Here is how to choose between them, written by the option with the least to gain from confusing you.

The three categories at a glance

OpenAI's Deployment Company launched 11 May 2026 at a $14bn valuation, with $4bn from TPG, Goldman Sachs, SoftBank, Capgemini and McKinsey, and Tomoro as its founding acquisition.2,5 Anthropic's enterprise services joint venture with Blackstone, Hellman & Friedman and Goldman Sachs landed in March 2026 at $1.5bn, alongside the Accenture and EPAM partnerships.3,4 Independents like QuantSpark are the third category.

Model lock-in

OpenAI DeployCoOpenAI models, by design
Anthropic JVClaude-first, by design
QuantSparkModel-agnostic. Right model per problem

IP ownership

OpenAI DeployCoNegotiated per engagement
Anthropic JVNegotiated per engagement
QuantSparkClient owns the code, data and weights

Pricing transparency

OpenAI DeployCoReported floors above five million
Anthropic JVEnterprise rate card, not published
QuantSpark£25k AiRE Diagnostic, £40k Labs Discovery, builds from £150k

Sector depth

OpenAI DeployCoFrontier brands, broad sweep
Anthropic JVBig Four scale through Accenture and EPAM
QuantSparkPE portcos, mid-market FS, UK public sector

Embedded headcount

OpenAI DeployCoc.150 FDEs ex-Tomoro, scaling fast
Anthropic JVc.30k Accenture staff retrained on Claude
QuantSparkc.70 senior practitioners, low ratio per client

What they will not do

OpenAI DeployCoRecommend a non-OpenAI model
Anthropic JVRecommend a non-Anthropic model
QuantSparkSell you a model. We sell outcomes

When to choose each

We will not pretend the labs lose every shoot-out. They do not. Here is the honest read on when each option wins.

OpenAI's DeployCo

When to choose the OpenAI Deployment Company

  • You need first-party access to OpenAI's frontier roadmap, including pre-release weights and reserved capacity.
  • Your business case absorbs a reported floor above five million dollars per engagement.
  • You are comfortable with OpenAI being both the model vendor and the implementation partner. The conflict is acceptable to you.
  • You want a single line of attack, and you trust OpenAI to make it a success.

They will deliver. Just be honest about the lock-in.

Anthropic's Enterprise JV

When to choose the Anthropic JV

  • You have already standardised on Claude across your business and want deep specialist integration.
  • You need Accenture or EPAM scale: thousands of bodies, global delivery centres, SAP and ServiceNow plumbing already in place.
  • You weight safety-aligned posture and constitutional AI more heavily than model performance on a given task.
  • You can absorb Big Four day rates and a Big Four delivery cadence.

Right partner if Claude is your standard and scale is the constraint.

Independent: QuantSpark

When to choose an independent forward-deployed partner

  • You want the best model for each problem, evaluated head-to-head, not the model your partner sells.
  • Your envelope is mid-market: £100k to £2m per programme, not five million plus.
  • You need IP-clean delivery. The code, the data pipelines and the weights stay with you.
  • You operate in private equity, financial services or UK public sector and want a partner who knows the regulatory surface.
  • You are willing to put GPT-5, Claude and an open-weights baseline head-to-head in week one and ship the winner.

This is our wedge. Model-neutral, sector-deep, transparently priced. No vendor partnerships clouding our advice.

How we actually build

Not every problem needs an agent

The right partner tells you when full autonomy adds cost and latency without improving the outcome, and reaches for the simplest thing that works.

Workflows

Predefined code paths, where the steps are known in advance. Well-instrumented workflows are where most enterprise value comes from today. They are predictable, cheap to run and easy to audit.

Agents

Systems where the model directs its own steps and tool use. The distinction is Anthropic’s, from “Building Effective AI Agents”. Agents earn their place only when the path cannot be fixed in advance and the extra cost buys a better result.

We embed and ship. We do not hand over a deck.

Forward-deployed means our engineers work inside your environment, against your data and constraints, and own the path from discovery to daily use. There is no advisor-to-builder handoff. A typical engagement runs like this.

1

Weeks 1 to 2

Discovery in the codebase

One or two engineers embed on your stack and work against your real data and constraints, not a sanitised sandbox.

2

Weeks 2 to 4

First working slice

A narrow use case runs end to end early, so value is demonstrated in software rather than promised in a deck.

3

Weeks 4 to 8

Integrate and instrument

We wire it into the systems your teams already use and add the evals, monitoring and controls that make it safe to rely on.

4

From week 8

Harden and hand to run

The workflow is hardened, documented and handed to an owner on your side. No advisor-to-builder handoff, no dependency on us to keep it alive.

This is the delivery method behind AiRE, our governed rollout engine.

See how AiRE takes agents to production

A passing test is not a working agent

Agents choose their own path, so several reasonable decisions can still reach the wrong outcome. We treat repeatability as a first-class metric: the same scenario run many times, with varied phrasing and injected tool failures, and every production incident turned into a permanent regression test. Autonomy is earned by demonstrated reliability and bounded by the cost of being wrong.

50%

VentureBeat, VB Pulse, June 2026

of enterprises surveyed had deployed an agent that passed internal evaluations yet still caused a customer-facing failure.

5%

VentureBeat, VB Pulse, June 2026

fully trust automated evaluations, which is why we design evals to expose failure, not to certify success.

The technology is the smaller half of the problem

Bolt an agent onto a process designed for people and you amplify its flaws. The harder half is redesigning the operating model around it: decision rights, exception handling, stop mechanisms and the new supervisory roles that agents create. Deloitte’s image is a jet engine on a bicycle; the engine is not the problem, the bicycle is.

We run change management and enablement in the same engagement as the build, not as a follow-on project. More than a decade and over 100 engagements, from FTSE 100 to government, taught us that value is won or lost in adoption, not in the model.

The independent test

Five questions to ask any forward-deployed partner

Self-qualifying. Run them past us and past every other shortlisted vendor. The answers tell you who you are actually hiring.

  1. 1

    Which model will you put in production for me, and why?

    If the answer is always the same model, you are buying lock-in, not advice.

  2. 2

    Who owns the code, data and fine-tuned weights when the engagement ends?

    If the answer is anything less than "you do", read the contract again.

  3. 3

    What is your minimum engagement size, and what does the first 30 days cost?

    A partner with nothing to hide will quote a paid diagnostic in writing.

  4. 4

    Will you publish a head-to-head model evaluation in week one?

    Independents say yes. Aligned partners find reasons it is not needed.

  5. 5

    Can you walk us through three engagements where you killed a project early?

    Honest forward-deployed teams have a graveyard. Sales-led ones do not.

The work behind the claim

Sector depth you can read for yourself

We say PE portcos, mid-market financial services and UK public sector. Here is the delivered evidence, sector by sector.

Sector

Private Equity

See the work

Sector

Financial Services

See the work

Sector

Public Sector

See the work

Put us through the test

The AiRE Diagnostic is two to four weeks to defined AI value and a clear AI plan. It's an understanding of where your model or processes are exposed or vulnerable, and a short list of AI bets ranked by ROI, feasibility and speed. It provides an honest call: roll out, build, buy, or wait. When the answer is build, QuantSpark Labs takes over.