An operator running a mid-market company is being pitched something new this quarter. The pitch combines model access, workflow redesign, and engineers who work inside the company for long enough to become part of how a function operates. The thing being sold is the operating layer of that function, on a multi-year arrangement, with a vendor and a services entity on the other side of the contract.
In the first two weeks of May, two leading frontier model vendors moved into that pitch directly. Anthropic announced a new AI services company with Blackstone, Hellman and Friedman, and Goldman Sachs Asset Management on May 4. Press reports put the venture at about 1.5 billion dollars. OpenAI launched the OpenAI Deployment Company on May 11 with more than 4 billion dollars of initial investment and 19 partner organizations across investment firms, consultancies, and systems integrators. OpenAI also agreed to acquire Tomoro, which would add roughly 150 Forward Deployed Engineers and deployment specialists once the deal closes. PwC and Anthropic expanded their alliance on May 14, with PwC rolling out Claude Code and Cowork starting with US teams, opening a joint Center of Excellence, and training 30,000 PwC US professionals on the platform.
Each of these is a different legal structure. OpenAI’s vehicle is majority owned and controlled by OpenAI. Anthropic’s vehicle is a standalone services firm with anchor capital from Blackstone, Hellman and Friedman, and Goldman Sachs Asset Management. PwC’s arrangement is a strategic alliance. Across all three, the buyer experiences the same combined offer: model access, services capacity, embedded engineering, and in the OpenAI and Anthropic cases, external capital scaling the deployment layer.
Forward Deployed Engineers are the mechanism that powers all of this. Palantir built the pattern for customers whose workflows changed faster than any outside vendor could track. Engineers go inside the company, learn how the work actually moves, and build production systems around what they learn. Over a sustained engagement, a meaningful share of the operating memory of that function ends up living in those engineers. They report to the services entity. The company carries an operating dependency on people who aren’t its employees and whose institutional knowledge of the function exceeds what most of the company’s own people have.
Consulting firms are already part of this. McKinsey and Company, Bain and Company, and Capgemini are among OpenAI DeployCo’s investment and integration partners. PwC has its own alliance with Anthropic on similar delivery terms. Accenture committed 3 billion dollars to its Data and AI practice in 2023, and announced a Forward Deployed Engineering program with ServiceNow on May 6 of this year. A McKinsey or PwC partner pitching enterprise AI work in 2026 is, increasingly, pitching work that runs on a model vendor’s stack. The vendor sits one layer up and is positioned to capture more of the underlying model economics behind each deployment.
The same pattern is emerging in the Gulf. G42 and Publicis Sapient signed an MOU earlier this year to explore a joint venture delivering AI first services across the UAE and what G42 calls the Global South, subject to final agreements and targeted for mid 2026. The MOU is early. The structural template is being copied regionally faster than most operators in the GCC have priced in.
For some functions, taking this offer is straightforwardly worth it. PwC reports insurance underwriting cycles compressed from ten weeks to ten days in Claude enabled deployments, alongside delivery improvements of up to 70 percent. That kind of compression can justify a more complicated procurement structure. The operating risks still have to be negotiated separately.
Worth working through, before signing.
Who owns the redesign of the function when the contract ends. The engagement produces process maps, permissions structures, and a way the work flows. That redesign is real intellectual property. When the engagement ends, the redesign has to stay with the company in usable form.
What the engagement exposes back to the vendor. Both major vendors publish commercial policies stating they do not train their models on enterprise inputs or outputs by default. That covers the training data layer. The operational layer is where the harder questions sit: what the embedded engineers see and remember, what logs and evals and deployment playbooks feed back to the services entity, what the vendor learns across its mid market clients in your sector that compounds into a product roadmap shaping how you’ll be served later.
What the exit costs. Swapping the underlying model vendor later in the engagement usually means months of operational pain followed by a period of partial functionality. The cloud migrations of the 2010s went worst for the companies that didn’t price exit cost into the original procurement.
Which of your own people grow during the engagement, and which atrophy. The Forward Deployed Engineer model assumes the customer’s people absorb the system over time. Often they absorb the outputs and not the structure. That capability gap is itself a kind of lock in, and it lives on the company’s side of the relationship.
How the rest of the company reads the engagement. The executive who signs sees workflow compression and operating tempo gains. The team using the system sees engineers from outside the company with embedded access to how the work runs. Both reads are accurate at the same time. Naming the asymmetry early makes the deployment land better than letting it surface later in the engagement.
What your competitors are doing on the same stack. The PE backed deployment companies will use their portfolio companies as proving grounds for whichever sectors they reach first. If the operators you compete with most directly are running the same vendor’s deployment company across the same set of patterns, the operating differentiation you used to carry gets thinner over time. An operator’s call is whether the engagement preserves enough idiosyncrasy in how the company actually runs.
The procurement question this year is which parts of an operator’s operating layer are worth handing to a vendor backed services entity. Some functions clear easily on the compression alone. Others clear only after a careful read of the contract, the data exposure, the exit cost, and the people implications. A lot of procurement conversations this quarter will skip the careful read, which is the version of the choice that costs more than it looks like it costs.