AI Is Starting to Write the Sales Proposals Humans Used to Spend Days Building

A request for quotation lands with an industrial supplier. The customer wants a technically complex product, the specification runs across hundreds or thousands of pages, and somebody has to work out whether the company can build it, which configuration fits, what it will cost and what the sales team can promise.

Sales engineers often sit in the middle of that work. They know enough about the product to understand the technical requirements and enough about the customer to turn those requirements into an offer. They also spend large parts of their week looking through documents, checking previous bids, asking colleagues for missing information and moving data between systems.

Munich-based Atira is trying to hand some of that work to AI agents. Founded in 2024 by former Palantir manager Florian Diegruber and former Google software engineer August DuMont Schütte, the company has built software that reads customer requests, extracts requirements and prepares technical documentation, configurations and pricing options. ABB E-Mobility is among its customers. Atira has raised $15 million in a round led by Accel and plans to more than double its team from 22 to around 50 people.

The software connects with systems such as SAP and Salesforce rather than asking companies to build another database beside them. Information generated or processed by the agent can go back into the original business system. Atira also sends its own engineers into customers to make the software work with existing data and processes, borrowing the forward-deployed engineering model associated with Palantir.

That last part says as much about enterprise AI as the agent itself.

The proposal was never just a document

A complex industrial offer draws information from several parts of a company. Sales knows what the customer asked for. Engineering knows which specifications are feasible. Pricing teams know the commercial limits. Previous projects contain answers that may already have taken days to establish.

Much of the work happens before anyone writes the final proposal.

A sales engineer may receive a specification requiring a particular voltage range, certification standard and installation environment. The company sells dozens of possible configurations. One fits technically but has a long delivery time. Another needs an engineering modification. A previous customer ordered something similar, but the details sit inside an old proposal rather than the product catalogue.

An AI agent reading those documents has a more demanding job than generating sales copy. It needs to identify the requirement, find the corresponding product information and recognise where the available data does not support an answer.

Atira concentrates on that gap. Its software processes requests and proposes configurations and prices while working with information already held in enterprise systems. Industrial companies that once ignored some tenders because they lacked the staff to prepare them could use the software to handle more requests without adding the same amount of manual work. One early investor describes that capacity as part of the commercial case for the product.

The attraction is easy to understand. The risk starts when the agent reaches a conclusion that looks plausible but is technically wrong.

A wrong sentence can become an expensive promise

Generative AI has spent the past few years getting better at producing convincing language. Industrial sales needs something less forgiving.

Suppose an agent reads a tender for charging infrastructure and selects a configuration that satisfies most of the customer’s requirements. One component does not meet a local certification rule. A salesperson reviewing a 70-page proposal does not notice. The customer awards the contract.

The problem now belongs to engineering, procurement and possibly legal. Companies therefore have to decide which parts of the proposal process the agent performs alone and where a person signs off. Extracting a voltage specification from a document carries a different risk from deciding that an existing product complies with it. Drafting a paragraph from verified data is different again from calculating a price or promising delivery.

The dividing line will vary by product. A company selling standardised components has fewer judgement calls than a manufacturer producing equipment around each customer’s site, regulation and technical constraints.

The systems also need a way to admit that the answer is missing. An agent that returns “I cannot find an approved configuration” is more useful than one that fills the gap with the statistically likely answer.

Old enterprise data becomes part of the sales process

Industrial companies already have much of the information these agents need. Finding it is another matter. Product specifications may sit in an ERP system, customer history in Salesforce, technical drawings in another application and pricing in spreadsheets maintained by a commercial team. Previous bids contain decisions that never made their way into structured data at all.

An agent working across that material exposes data problems that employees previously worked around manually. If product names differ between SAP and the CRM, someone has to reconcile them. If the approved price list is unclear, the agent cannot safely choose one. If a product change never reached the old sales documentation, search retrieves obsolete information alongside the current version.

AI therefore pushes companies back into integration work. Atira’s decision to use forward-deployed engineers reflects that reality. Palantir made the role well known by sending technical staff into customers to connect software with the organisation’s actual data and processes rather than expecting a standard installation to handle everything. Enterprise AI companies are adopting the same model because many failures appear only after software meets the customer’s systems.

SAP is heading in a similar direction from inside the established software stack. Its Joule Studio lets companies build agents around existing business processes, including supply-chain work. The appeal is obvious for companies already running SAP because the agent starts close to the transactions and data it needs.

Sales engineering is a good test for enterprise agents

Many early corporate AI tools went after work with low consequences when the output was wrong. Summarising a meeting, drafting an internal email or preparing the first version of a presentation saved time without giving the software much authority.

Sales engineering sits closer to revenue and contractual commitments. An agent working there touches product knowledge, commercial decisions and customer promises. It also deals with work that companies can measure. A bid took twelve days before deployment and four days afterwards. Engineers answered 80 questions manually and now answer 20. The company submitted tenders it previously declined because the team lacked time.

Those numbers give enterprise buyers something stronger than a demonstration of fluent text. They also make poor implementations easier to spot. If engineers spend the saved hours checking everything the agent produced, the automation has moved the work rather than removed it. If the system drafts more proposals but creates more corrections later in the sales cycle, throughput alone says little.

Atira’s customer base has grown to more than a dozen companies in its first year, ranging from Mittelstand businesses to ABB E-Mobility. Investors have backed the company partly because customers are already expanding usage rather than keeping the software inside small pilot projects.

The salesperson does not disappear from the deal

Industrial sales involves information that software handles well and negotiations that remain harder to reduce to a workflow. A customer may ask for a lower price because a competitor has entered the tender. An engineer may know that production has had trouble with the configuration the agent selected. A salesperson may realise that a formally minor technical requirement matters disproportionately to the person making the purchasing decision.

None of that makes automation irrelevant. It changes where people spend their time. If an agent extracts 300 technical requirements, finds the corresponding product documentation and prepares a first configuration, an engineer can spend the review on the unusual 20 rather than manually processing all 300. The salesperson reaches the customer sooner. Specialists receive questions that require judgement instead of requests to find information stored elsewhere in the company.

The software becomes more ambitious when it starts deciding rather than preparing. Companies will have to make those permissions explicit. An agent might draft a price while finance approves it. It might recommend a product configuration while an engineer accepts it. Routine offers may eventually need less review than unusual tenders.

The safest boundary will move as companies gain evidence from real proposals rather than demonstrations.

The next AI battle is inside existing work

Atira is not alone in going after industrial sales. Berlin-based Plato has raised $14.5 million to automate work for wholesalers, including quotation and order preparation, and SAP itself wants agents to automate more of the business processes already running through its software.

The competition will not be decided by which company has the most impressive chatbot. Industrial customers need software that knows where approved data sits, understands which system owns which number and records the final decision in the place employees already use. An agent that generates a technically correct proposal but leaves the CRM, ERP and pricing system out of sync creates another job for somebody to clean up.

Atira’s engineers currently go into customer organisations to solve those problems directly. As the company grows from 22 employees towards roughly 50, it will have to show how much of that work it can standardise without losing the close integration that helped win its first customers.

For ABB and other industrial buyers, the more immediate test sits with the next request for quotation. If the sales engineer opens it and spends less time searching old files, copying specifications and rebuilding work the company has done before, the agent has already taken over part of the job. The engineer still decides what the company is prepared to promise.