One way to phrase your question is asking if the rate for the transition $A^*\rightarrow A+\gamma$ (where $A^*$ is some excited state of the atom $A$, taken to be neutral for simplicity to avoid similar issues in QED) should actually be zero since, actually, there always should be gravitons around, corresponding instead to the following processes $A^*\rightarrow A+\gamma+g$, $A^*\rightarrow A+\gamma+2g,\ldots$.
It turns out that the answer to this is affirmative, because of a well-known property of gravity associated with the graviton being massless and hence producing infrared (IR) singularities. They require the inclusion of an IR regulator $\lambda$ (to be chosen smaller than the smallest physical mass scale in the problem) an that the rate for any exclusive process such as the one of your question $A^*\rightarrow A+\gamma$ actually vanishes when the IR regulator is removed, $\lambda\rightarrow 0$. Moreover, the rate that is instead finite is the same one you wanted but with gravitational bremsstrahlung (that is emission of infinitely many soft gravitons of total finite energy) added in the initial and final state. The classical limit of such a process corresponds to the emission of classical gravitational bremsstrahlung.
More explicitly, S. Weinberg (see e.g. vol.I chapter 13.2 and original references therein), generalizing earlier work for a similar phenomenon with photon-bremsstrahlung in QED by Yiennie, Fratsci and Suura, has shown that the rate $\Gamma^\lambda_{\alpha\rightarrow \beta}$ for any process $\alpha\rightarrow \beta$ (with or without gravitons) involving at most gravitons (real or virtual) with energy $E$ above the IR regulator , $\lambda<E$, vanishes as
$$
\Gamma^{\lambda}_{\alpha\rightarrow\beta}=\left(\frac{\lambda}{\Lambda}\right)^A\Gamma^{\Lambda}_{\alpha\rightarrow\beta}\rightarrow 0\,,\qquad A=\frac{G}{2\pi}\sum_{n\,, m}\eta_n \eta_{m} m_n m_m\frac{1+\beta_{nm}^2}{\beta_{nm}\sqrt{1-\beta_{nm}^2}}\log\frac{1+\beta_{nm}}{1-\beta_{nm}}>0
$$
for $\lambda\rightarrow 0$, where $m_n$ is the mass of the $n$-th particle line, and $\eta_{n}=\pm 1$ depending whether such a particle is ingoing or outgoing, and $\beta_{nm}=\sqrt{1-m_n^2 m_m^2/(p_n p_m)^2}$ is the relative velocity between the pairs $n$ and $m$.
What is instead finite when the IR cutoff is removed is the rate $\Gamma^{\lambda;\, E}_{\alpha\rightarrow\beta}$ for the same process but with arbitrarily many soft gravitons emitted with energy larger than $\lambda$ and below some $E$, such that
$$
\Gamma^{\lambda;\, E}_{\alpha\rightarrow\beta} \rightarrow \left(\frac{E}{\lambda}\right)^{A}\Gamma^{\lambda}_{\alpha\rightarrow\beta}=\left(\frac{E}{\Lambda}\right)^A\Gamma^{\Lambda}_{\alpha\rightarrow\beta}
$$
up to a known overall o(1)-function that I am omitting for simplicity.
